sys/ofed/drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  * Copyright (c) 2016 Chelsio Communications.  All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <linux/completion.h>
  38 #include <linux/in.h>
  39 #include <linux/in6.h>
  40 #include <linux/mutex.h>
  41 #include <linux/random.h>
  42 #include <linux/idr.h>
  43 #include <linux/inetdevice.h>
  44 #include <linux/slab.h>
  45 #include <linux/module.h>
  46 #include <linux/string.h>
  47 #include <net/route.h>
  48
  49 #include <net/tcp.h>
  50 #include <net/ipv6.h>
  51
  52 #include <rdma/rdma_cm.h>
  53 #include <rdma/rdma_cm_ib.h>
  54 #include <rdma/ib_cache.h>
  55 #include <rdma/ib_cm.h>
  56 #include <rdma/ib_sa.h>
  57 #include <rdma/iw_cm.h>
  58
  59 MODULE_AUTHOR("Sean Hefty");
  60 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 #define CMA_CM_RESPONSE_TIMEOUT 20
  64 #define CMA_MAX_CM_RETRIES 15
  65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  66 #define CMA_IBOE_PACKET_LIFETIME 18
  67
  68 static int cma_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
  69 module_param_named(cma_response_timeout, cma_response_timeout, int, 0644);
  70 MODULE_PARM_DESC(cma_response_timeout, "CMA_CM_RESPONSE_TIMEOUT (default=20)");
  71
  72 static int def_prec2sl = 3;
  73 module_param_named(def_prec2sl, def_prec2sl, int, 0644);
  74 MODULE_PARM_DESC(def_prec2sl, "Default value for SL priority with RoCE. Valid values 0 - 7");
  75
  76 static int unify_tcp_port_space = 1;
  77 module_param(unify_tcp_port_space, int, 0644);
  78 MODULE_PARM_DESC(unify_tcp_port_space, "Unify the host TCP and RDMA port "
  79                  "space allocation (default=1)");
  80
  81 static int debug_level = 0;
  82 #define cma_pr(level, priv, format, arg...)             \
  83         printk(level "CMA: %p: %s: " format, ((struct rdma_id_priv *) priv) , __func__, ## arg)
  84
  85 #define cma_dbg(priv, format, arg...)           \
  86         do { if (debug_level) cma_pr(KERN_DEBUG, priv, format, ## arg); } while (0)
  87
  88 #define cma_warn(priv, format, arg...)          \
  89         cma_pr(KERN_WARNING, priv, format, ## arg)
  90
  91 #define CMA_GID_FMT        "%2.2x%2.2x:%2.2x%2.2x"
  92 #define CMA_GID_RAW_ARG(gid) ((u8 *)(gid))[12],\
  93                                    ((u8 *)(gid))[13],\
  94                                    ((u8 *)(gid))[14],\
  95                                    ((u8 *)(gid))[15]
  96
  97 #define CMA_GID_ARG(gid)   CMA_GID_RAW_ARG((gid).raw)
  98 #define cma_debug_path(priv, pfx, p) \
  99         cma_dbg(priv, pfx "sgid=" CMA_GID_FMT ",dgid="  \
 100                 CMA_GID_FMT "\n", CMA_GID_ARG(p.sgid),  \
 101                 CMA_GID_ARG(p.dgid))
 102
 103 #define cma_debug_gid(priv, g) \
 104         cma_dbg(priv, "gid=" CMA_GID_FMT "\n", CMA_GID_ARG(g)
 105
 106 module_param_named(debug_level, debug_level, int, 0644);
 107 MODULE_PARM_DESC(debug_level, "debug level default=0");
 108
 109 static void cma_add_one(struct ib_device *device);
 110 static void cma_remove_one(struct ib_device *device);
 111
 112 static struct ib_client cma_client = {
 113         .name   = "cma",
 114         .add    = cma_add_one,
 115         .remove = cma_remove_one
 116 };
 117
 118 static struct ib_sa_client sa_client;
 119 static struct rdma_addr_client addr_client;
 120 static LIST_HEAD(dev_list);
 121 static LIST_HEAD(listen_any_list);
 122 static DEFINE_MUTEX(lock);
 123 static struct workqueue_struct *cma_wq;
 124 static struct workqueue_struct *cma_free_wq;
 125 static DEFINE_IDR(sdp_ps);
 126 static DEFINE_IDR(tcp_ps);
 127 static DEFINE_IDR(udp_ps);
 128 static DEFINE_IDR(ipoib_ps);
 129 static DEFINE_IDR(ib_ps);
 130
 131 struct cma_device {
 132         struct list_head        list;
 133         struct ib_device        *device;
 134         struct completion       comp;
 135         atomic_t                refcount;
 136         struct list_head        id_list;
 137 };
 138
 139 struct rdma_bind_list {
 140         struct idr              *ps;
 141         struct hlist_head       owners;
 142         unsigned short          port;
 143 };
 144
 145 enum {
 146         CMA_OPTION_AFONLY,
 147 };
 148
 149 /*
 150  * Device removal can occur at anytime, so we need extra handling to
 151  * serialize notifying the user of device removal with other callbacks.
 152  * We do this by disabling removal notification while a callback is in process,
 153  * and reporting it after the callback completes.
 154  */
 155 struct rdma_id_private {
 156         struct rdma_cm_id       id;
 157
 158         struct rdma_bind_list   *bind_list;
 159         struct socket           *sock;
 160         struct hlist_node       node;
 161         struct list_head        list; /* listen_any_list or cma_device.list */
 162         struct list_head        listen_list; /* per device listens */
 163         struct cma_device       *cma_dev;
 164         struct list_head        mc_list;
 165
 166         int                     internal_id;
 167         enum rdma_cm_state      state;
 168         spinlock_t              lock;
 169         spinlock_t              cm_lock;
 170         struct mutex            qp_mutex;
 171
 172         struct completion       comp;
 173         atomic_t                refcount;
 174         struct mutex            handler_mutex;
 175         struct work_struct      work;  /* garbage coll */
 176
 177         int                     backlog;
 178         int                     timeout_ms;
 179         struct ib_sa_query      *query;
 180         int                     query_id;
 181         union {
 182                 struct ib_cm_id *ib;
 183                 struct iw_cm_id *iw;
 184         } cm_id;
 185
 186         u32                     seq_num;
 187         u32                     qkey;
 188         u32                     qp_num;
 189         pid_t                   owner;
 190         u32                     options;
 191         u8                      srq;
 192         u8                      tos;
 193         u8                      reuseaddr;
 194         u8                      afonly;
 195         int                     qp_timeout;
 196         /* cache for mc record params */
 197         struct ib_sa_mcmember_rec rec;
 198         int is_valid_rec;
 199 };
 200
 201 struct cma_multicast {
 202         struct rdma_id_private *id_priv;
 203         union {
 204                 struct ib_sa_multicast *ib;
 205         } multicast;
 206         struct list_head        list;
 207         void                    *context;
 208         struct sockaddr_storage addr;
 209         struct kref             mcref;
 210 };
 211
 212 struct cma_work {
 213         struct work_struct      work;
 214         struct rdma_id_private  *id;
 215         enum rdma_cm_state      old_state;
 216         enum rdma_cm_state      new_state;
 217         struct rdma_cm_event    event;
 218 };
 219
 220 struct cma_ndev_work {
 221         struct work_struct      work;
 222         struct rdma_id_private  *id;
 223         struct rdma_cm_event    event;
 224 };
 225
 226 struct iboe_mcast_work {
 227         struct work_struct       work;
 228         struct rdma_id_private  *id;
 229         struct cma_multicast    *mc;
 230 };
 231
 232 union cma_ip_addr {
 233         struct in6_addr ip6;
 234         struct {
 235                 __be32 pad[3];
 236                 __be32 addr;
 237         } ip4;
 238 };
 239
 240 struct cma_hdr {
 241         u8 cma_version;
 242         u8 ip_version;  /* IP version: 7:4 */
 243         __be16 port;
 244         union cma_ip_addr src_addr;
 245         union cma_ip_addr dst_addr;
 246 };
 247
 248 struct sdp_hh {
 249         u8 bsdh[16];
 250         u8 sdp_version; /* Major version: 7:4 */
 251         u8 ip_version;  /* IP version: 7:4 */
 252         u8 sdp_specific1[10];
 253         __be16 port;
 254         __be16 sdp_specific2;
 255         union cma_ip_addr src_addr;
 256         union cma_ip_addr dst_addr;
 257 };
 258
 259 struct sdp_hah {
 260         u8 bsdh[16];
 261         u8 sdp_version;
 262 };
 263
 264 #define CMA_VERSION 0x00
 265 #define SDP_MAJ_VERSION 0x2
 266
 267 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 268 {
 269         unsigned long flags;
 270         int ret;
 271
 272         spin_lock_irqsave(&id_priv->lock, flags);
 273         ret = (id_priv->state == comp);
 274         spin_unlock_irqrestore(&id_priv->lock, flags);
 275         return ret;
 276 }
 277
 278 static int cma_comp_exch(struct rdma_id_private *id_priv,
 279                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 280 {
 281         unsigned long flags;
 282         int ret;
 283
 284         spin_lock_irqsave(&id_priv->lock, flags);
 285         if ((ret = (id_priv->state == comp)))
 286                 id_priv->state = exch;
 287         spin_unlock_irqrestore(&id_priv->lock, flags);
 288         return ret;
 289 }
 290
 291 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 292                                    enum rdma_cm_state exch)
 293 {
 294         unsigned long flags;
 295         enum rdma_cm_state old;
 296
 297         spin_lock_irqsave(&id_priv->lock, flags);
 298         old = id_priv->state;
 299         id_priv->state = exch;
 300         spin_unlock_irqrestore(&id_priv->lock, flags);
 301         return old;
 302 }
 303
 304 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 305 {
 306         return hdr->ip_version >> 4;
 307 }
 308
 309 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 310 {
 311         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 312 }
 313
 314 static inline u8 sdp_get_majv(u8 sdp_version)
 315 {
 316         return sdp_version >> 4;
 317 }
 318
 319 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 320 {
 321         return hh->ip_version >> 4;
 322 }
 323
 324 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 325 {
 326         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 327 }
 328
 329 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 330                               struct cma_device *cma_dev)
 331 {
 332         atomic_inc(&cma_dev->refcount);
 333         id_priv->cma_dev = cma_dev;
 334         id_priv->id.device = cma_dev->device;
 335         id_priv->id.route.addr.dev_addr.transport =
 336                 rdma_node_get_transport(cma_dev->device->node_type);
 337         list_add_tail(&id_priv->list, &cma_dev->id_list);
 338 }
 339
 340 static inline void cma_deref_dev(struct cma_device *cma_dev)
 341 {
 342         if (atomic_dec_and_test(&cma_dev->refcount))
 343                 complete(&cma_dev->comp);
 344 }
 345
 346 static inline void release_mc(struct kref *kref)
 347 {
 348         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 349
 350         kfree(mc->multicast.ib);
 351         kfree(mc);
 352 }
 353
 354 static void cma_release_dev(struct rdma_id_private *id_priv)
 355 {
 356         mutex_lock(&lock);
 357         list_del(&id_priv->list);
 358         cma_deref_dev(id_priv->cma_dev);
 359         id_priv->cma_dev = NULL;
 360         mutex_unlock(&lock);
 361 }
 362
 363 static int cma_set_qkey(struct rdma_id_private *id_priv)
 364 {
 365         struct ib_sa_mcmember_rec rec;
 366         int ret = 0;
 367
 368         if (id_priv->qkey)
 369                 return 0;
 370
 371         switch (id_priv->id.ps) {
 372         case RDMA_PS_UDP:
 373                 id_priv->qkey = RDMA_UDP_QKEY;
 374                 break;
 375         case RDMA_PS_IPOIB:
 376                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 377                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 378                                              id_priv->id.port_num, &rec.mgid,
 379                                              &rec);
 380                 if (!ret)
 381                         id_priv->qkey = be32_to_cpu(rec.qkey);
 382                 break;
 383         default:
 384                 break;
 385         }
 386         return ret;
 387 }
 388
 389 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 390 {
 391         int i;
 392         int err;
 393         struct ib_port_attr props;
 394         union ib_gid tmp;
 395
 396         err = ib_query_port(device, port_num, &props);
 397         if (err)
 398                 return 1;
 399
 400         for (i = 0; i < props.gid_tbl_len; ++i) {
 401                 err = ib_query_gid(device, port_num, i, &tmp);
 402                 if (err)
 403                         return 1;
 404                 if (!memcmp(&tmp, gid, sizeof tmp))
 405                         return 0;
 406         }
 407
 408         return -EAGAIN;
 409 }
 410
 411 int
 412 rdma_find_cmid_laddr(struct sockaddr_in *local_addr, unsigned short dev_type,
 413                                                         void **cm_id)
 414 {
 415         int ret;
 416         u8 port;
 417         int found_dev = 0, found_cmid = 0;
 418         struct rdma_id_private  *id_priv;
 419         struct rdma_id_private  *dev_id_priv;
 420         struct cma_device       *cma_dev;
 421         struct rdma_dev_addr    dev_addr;
 422         union ib_gid            gid;
 423         enum rdma_link_layer dev_ll = dev_type == ARPHRD_INFINIBAND ?
 424                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 425
 426         memset(&dev_addr, 0, sizeof(dev_addr));
 427
 428         ret = rdma_translate_ip((struct sockaddr *)local_addr,
 429                                                         &dev_addr, NULL);
 430         if (ret)
 431                 goto err;
 432
 433         /* find rdma device based on MAC address/gid */
 434         mutex_lock(&lock);
 435
 436         memcpy(&gid, dev_addr.src_dev_addr +
 437                rdma_addr_gid_offset(&dev_addr), sizeof(gid));
 438
 439         list_for_each_entry(cma_dev, &dev_list, list)
 440                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port)
 441                         if ((rdma_port_get_link_layer(cma_dev->device, port) ==
 442                                                                  dev_ll) &&
 443                          (rdma_node_get_transport(cma_dev->device->node_type) ==
 444                                                         RDMA_TRANSPORT_IWARP)) {
 445                                         ret = find_gid_port(cma_dev->device,
 446                                                                 &gid, port);
 447                                         if (!ret) {
 448                                                 found_dev = 1;
 449                                                 goto out;
 450                                         } else if (ret == 1) {
 451                                                 mutex_unlock(&lock);
 452                                                 goto err;
 453                                         }
 454                         }
 455 out:
 456         mutex_unlock(&lock);
 457
 458         if (!found_dev)
 459                 goto err;
 460
 461         /* Traverse through the list of listening cm_id's to find the
 462          * desired cm_id based on rdma device & port number.
 463          */
 464         list_for_each_entry(id_priv, &listen_any_list, list)
 465                 list_for_each_entry(dev_id_priv, &id_priv->listen_list,
 466                                                  listen_list)
 467                         if (dev_id_priv->cma_dev == cma_dev)
 468                                 if (dev_id_priv->cm_id.iw->local_addr.sin_port
 469                                                 == local_addr->sin_port) {
 470                                         *cm_id = (void *)dev_id_priv->cm_id.iw;
 471                                         found_cmid = 1;
 472                                 }
 473         return found_cmid ? 0 : -ENODEV;
 474
 475 err:
 476         return -ENODEV;
 477 }
 478 EXPORT_SYMBOL(rdma_find_cmid_laddr);
 479
 480 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 481 {
 482         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 483         struct cma_device *cma_dev;
 484         union ib_gid gid, iboe_gid;
 485         int ret = -ENODEV;
 486         u8 port;
 487         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 488                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 489
 490         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 491             id_priv->id.ps == RDMA_PS_IPOIB)
 492                 return -EINVAL;
 493
 494         mutex_lock(&lock);
 495         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 496                     &iboe_gid);
 497
 498         memcpy(&gid, dev_addr->src_dev_addr +
 499                rdma_addr_gid_offset(dev_addr), sizeof gid);
 500         list_for_each_entry(cma_dev, &dev_list, list) {
 501                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 502                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 503                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 504                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 505                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 506                                 else
 507                                         ret = find_gid_port(cma_dev->device, &gid, port);
 508
 509                                 if (!ret) {
 510                                         id_priv->id.port_num = port;
 511                                         goto out;
 512                                 } else if (ret == 1)
 513                         break;
 514         }
 515                 }
 516         }
 517
 518 out:
 519         if (!ret)
 520                 cma_attach_to_dev(id_priv, cma_dev);
 521
 522         mutex_unlock(&lock);
 523         return ret;
 524 }
 525
 526 static void cma_deref_id(struct rdma_id_private *id_priv)
 527 {
 528         if (atomic_dec_and_test(&id_priv->refcount))
 529                 complete(&id_priv->comp);
 530 }
 531
 532 static int cma_disable_callback(struct rdma_id_private *id_priv,
 533                                 enum rdma_cm_state state)
 534 {
 535         mutex_lock(&id_priv->handler_mutex);
 536         if (id_priv->state != state) {
 537                 mutex_unlock(&id_priv->handler_mutex);
 538                 return -EINVAL;
 539         }
 540         return 0;
 541 }
 542
 543 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 544                                   void *context, enum rdma_port_space ps,
 545                                   enum ib_qp_type qp_type)
 546 {
 547         struct rdma_id_private *id_priv;
 548
 549         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 550         if (!id_priv)
 551                 return ERR_PTR(-ENOMEM);
 552
 553         id_priv->owner = curthread->td_proc->p_pid;
 554         id_priv->state = RDMA_CM_IDLE;
 555         id_priv->id.context = context;
 556         id_priv->id.event_handler = event_handler;
 557         id_priv->id.ps = ps;
 558         id_priv->id.qp_type = qp_type;
 559         spin_lock_init(&id_priv->lock);
 560         spin_lock_init(&id_priv->cm_lock);
 561         mutex_init(&id_priv->qp_mutex);
 562         init_completion(&id_priv->comp);
 563         atomic_set(&id_priv->refcount, 1);
 564         mutex_init(&id_priv->handler_mutex);
 565         INIT_LIST_HEAD(&id_priv->listen_list);
 566         INIT_LIST_HEAD(&id_priv->mc_list);
 567         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 568
 569         return &id_priv->id;
 570 }
 571 EXPORT_SYMBOL(rdma_create_id);
 572
 573 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 574 {
 575         struct ib_qp_attr qp_attr;
 576         int qp_attr_mask, ret;
 577
 578         qp_attr.qp_state = IB_QPS_INIT;
 579         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 580         if (ret)
 581                 return ret;
 582
 583         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 584         if (ret)
 585                 return ret;
 586
 587         qp_attr.qp_state = IB_QPS_RTR;
 588         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 589         if (ret)
 590                 return ret;
 591
 592         qp_attr.qp_state = IB_QPS_RTS;
 593         qp_attr.sq_psn = 0;
 594         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 595
 596         return ret;
 597 }
 598
 599 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 600 {
 601         struct ib_qp_attr qp_attr;
 602         int qp_attr_mask, ret;
 603
 604         qp_attr.qp_state = IB_QPS_INIT;
 605         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 606         if (ret)
 607                 return ret;
 608
 609         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 610 }
 611
 612 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 613                    struct ib_qp_init_attr *qp_init_attr)
 614 {
 615         struct rdma_id_private *id_priv;
 616         struct ib_qp *qp;
 617         int ret;
 618
 619         id_priv = container_of(id, struct rdma_id_private, id);
 620         if (id->device != pd->device)
 621                 return -EINVAL;
 622
 623         qp = ib_create_qp(pd, qp_init_attr);
 624         if (IS_ERR(qp))
 625                 return PTR_ERR(qp);
 626
 627         if (id->qp_type == IB_QPT_UD)
 628                 ret = cma_init_ud_qp(id_priv, qp);
 629         else
 630                 ret = cma_init_conn_qp(id_priv, qp);
 631         if (ret)
 632                 goto err;
 633
 634         id->qp = qp;
 635         id_priv->qp_num = qp->qp_num;
 636         id_priv->srq = (qp->srq != NULL);
 637         return 0;
 638 err:
 639         ib_destroy_qp(qp);
 640         return ret;
 641 }
 642 EXPORT_SYMBOL(rdma_create_qp);
 643
 644 void rdma_destroy_qp(struct rdma_cm_id *id)
 645 {
 646         struct rdma_id_private *id_priv;
 647
 648         id_priv = container_of(id, struct rdma_id_private, id);
 649         mutex_lock(&id_priv->qp_mutex);
 650         ib_destroy_qp(id_priv->id.qp);
 651         id_priv->id.qp = NULL;
 652         mutex_unlock(&id_priv->qp_mutex);
 653 }
 654 EXPORT_SYMBOL(rdma_destroy_qp);
 655
 656 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 657                              struct rdma_conn_param *conn_param)
 658 {
 659         struct ib_qp_attr qp_attr;
 660         int qp_attr_mask, ret;
 661         union ib_gid sgid;
 662
 663         mutex_lock(&id_priv->qp_mutex);
 664         if (!id_priv->id.qp) {
 665                 ret = 0;
 666                 goto out;
 667         }
 668
 669         /* Need to update QP attributes from default values. */
 670         qp_attr.qp_state = IB_QPS_INIT;
 671         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 672         if (ret)
 673                 goto out;
 674
 675         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 676         if (ret)
 677                 goto out;
 678
 679         qp_attr.qp_state = IB_QPS_RTR;
 680         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 681         if (ret)
 682                 goto out;
 683         ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
 684                            qp_attr.ah_attr.grh.sgid_index, &sgid);
 685         if (ret)
 686                 goto out;
 687
 688         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type)
 689             == RDMA_TRANSPORT_IB &&
 690             rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)
 691             == IB_LINK_LAYER_ETHERNET) {
 692                 u32 scope_id = rdma_get_ipv6_scope_id(id_priv->id.device,
 693                     id_priv->id.port_num);
 694
 695                 ret = rdma_addr_find_smac_by_sgid(&sgid, qp_attr.smac, NULL,
 696                     scope_id);
 697                 if (ret)
 698                         goto out;
 699         }
 700
 701         if (conn_param)
 702                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 703         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 704 out:
 705         mutex_unlock(&id_priv->qp_mutex);
 706         return ret;
 707 }
 708
 709 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 710                              struct rdma_conn_param *conn_param)
 711 {
 712         struct ib_qp_attr qp_attr;
 713         int qp_attr_mask, ret;
 714
 715         mutex_lock(&id_priv->qp_mutex);
 716         if (!id_priv->id.qp) {
 717                 ret = 0;
 718                 goto out;
 719         }
 720
 721         qp_attr.qp_state = IB_QPS_RTS;
 722         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 723         if (ret)
 724                 goto out;
 725
 726         if (conn_param)
 727                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 728
 729         if (id_priv->qp_timeout && id_priv->id.qp->qp_type == IB_QPT_RC) {
 730                 qp_attr.timeout = id_priv->qp_timeout;
 731                 qp_attr_mask |= IB_QP_TIMEOUT;
 732         }
 733
 734         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 735 out:
 736         mutex_unlock(&id_priv->qp_mutex);
 737         return ret;
 738 }
 739
 740 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 741 {
 742         struct ib_qp_attr qp_attr;
 743         int ret;
 744
 745         mutex_lock(&id_priv->qp_mutex);
 746         if (!id_priv->id.qp) {
 747                 ret = 0;
 748                 goto out;
 749         }
 750
 751         qp_attr.qp_state = IB_QPS_ERR;
 752         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 753 out:
 754         mutex_unlock(&id_priv->qp_mutex);
 755         return ret;
 756 }
 757
 758 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 759                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 760 {
 761         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 762         int ret;
 763         u16 pkey;
 764
 765         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 766             IB_LINK_LAYER_INFINIBAND)
 767                 pkey = ib_addr_get_pkey(dev_addr);
 768         else
 769                 pkey = 0xffff;
 770
 771         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 772                                   pkey, &qp_attr->pkey_index);
 773         if (ret)
 774                 return ret;
 775
 776         qp_attr->port_num = id_priv->id.port_num;
 777         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 778
 779         if (id_priv->id.qp_type == IB_QPT_UD) {
 780                 ret = cma_set_qkey(id_priv);
 781                 if (ret)
 782                         return ret;
 783
 784                 qp_attr->qkey = id_priv->qkey;
 785                 *qp_attr_mask |= IB_QP_QKEY;
 786         } else {
 787                 qp_attr->qp_access_flags = 0;
 788                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 789         }
 790         return 0;
 791 }
 792
 793 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 794                        int *qp_attr_mask)
 795 {
 796         struct rdma_id_private *id_priv;
 797         int ret = 0;
 798
 799         id_priv = container_of(id, struct rdma_id_private, id);
 800         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 801         case RDMA_TRANSPORT_IB:
 802                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 803                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 804                 else
 805                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 806                                                  qp_attr_mask);
 807                 if (qp_attr->qp_state == IB_QPS_RTR)
 808                         qp_attr->rq_psn = id_priv->seq_num;
 809                 break;
 810         case RDMA_TRANSPORT_IWARP:
 811         case RDMA_TRANSPORT_SCIF:
 812                 if (!id_priv->cm_id.iw) {
 813                         qp_attr->qp_access_flags = 0;
 814                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 815                 } else
 816                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 817                                                  qp_attr_mask);
 818                 break;
 819         default:
 820                 ret = -ENOSYS;
 821                 break;
 822         }
 823
 824         return ret;
 825 }
 826 EXPORT_SYMBOL(rdma_init_qp_attr);
 827
 828 static inline int cma_zero_addr(struct sockaddr *addr)
 829 {
 830         struct in6_addr *ip6;
 831
 832         if (addr->sa_family == AF_INET)
 833                 return ipv4_is_zeronet(
 834                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 835         else {
 836                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 837                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 838                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 839         }
 840 }
 841
 842 static inline int cma_loopback_addr(struct sockaddr *addr)
 843 {
 844         if (addr->sa_family == AF_INET)
 845                 return ipv4_is_loopback(
 846                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 847         else
 848                 return ipv6_addr_loopback(
 849                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 850 }
 851
 852 static inline int cma_any_addr(struct sockaddr *addr)
 853 {
 854         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 855 }
 856 int
 857 rdma_cma_any_addr(struct sockaddr *addr)
 858 {
 859         return cma_any_addr(addr);
 860 }
 861 EXPORT_SYMBOL(rdma_cma_any_addr);
 862
 863 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 864 {
 865         if (src->sa_family != dst->sa_family)
 866                 return -1;
 867
 868         switch (src->sa_family) {
 869         case AF_INET:
 870                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 871                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 872         default:
 873                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 874                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 875         }
 876 }
 877
 878 static inline __be16 cma_port(struct sockaddr *addr)
 879 {
 880         if (addr->sa_family == AF_INET)
 881                 return ((struct sockaddr_in *) addr)->sin_port;
 882         else
 883                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 884 }
 885
 886 static inline int cma_any_port(struct sockaddr *addr)
 887 {
 888         return !cma_port(addr);
 889 }
 890
 891 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 892                             u8 *ip_ver, __be16 *port,
 893                             union cma_ip_addr **src, union cma_ip_addr **dst)
 894 {
 895         switch (ps) {
 896         case RDMA_PS_SDP:
 897                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 898                     SDP_MAJ_VERSION)
 899                         return -EINVAL;
 900
 901                 *ip_ver = sdp_get_ip_ver(hdr);
 902                 *port   = ((struct sdp_hh *) hdr)->port;
 903                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 904                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 905                 break;
 906         default:
 907                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 908                         return -EINVAL;
 909
 910                 *ip_ver = cma_get_ip_ver(hdr);
 911                 *port   = ((struct cma_hdr *) hdr)->port;
 912                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 913                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 914                 break;
 915         }
 916
 917         if (*ip_ver != 4 && *ip_ver != 6)
 918                 return -EINVAL;
 919         return 0;
 920 }
 921
 922 static void cma_save_net_info(struct rdma_addr *addr,
 923                               struct rdma_addr *listen_addr,
 924                               u8 ip_ver, __be16 port,
 925                               union cma_ip_addr *src, union cma_ip_addr *dst)
 926 {
 927         struct sockaddr_in *listen4, *ip4;
 928         struct sockaddr_in6 *listen6, *ip6;
 929
 930         switch (ip_ver) {
 931         case 4:
 932                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 933                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 934                 ip4->sin_family = listen4->sin_family;
 935                 ip4->sin_addr.s_addr = dst->ip4.addr;
 936                 ip4->sin_port = listen4->sin_port;
 937                 ip4->sin_len = sizeof(struct sockaddr_in);
 938
 939                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 940                 ip4->sin_family = listen4->sin_family;
 941                 ip4->sin_addr.s_addr = src->ip4.addr;
 942                 ip4->sin_port = port;
 943                 ip4->sin_len = sizeof(struct sockaddr_in);
 944                 break;
 945         case 6:
 946                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 947                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 948                 ip6->sin6_family = listen6->sin6_family;
 949                 ip6->sin6_addr = dst->ip6;
 950                 ip6->sin6_port = listen6->sin6_port;
 951                 ip6->sin6_len = sizeof(struct sockaddr_in6);
 952                 ip6->sin6_scope_id = listen6->sin6_scope_id;
 953
 954                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 955                 ip6->sin6_family = listen6->sin6_family;
 956                 ip6->sin6_addr = src->ip6;
 957                 ip6->sin6_port = port;
 958                 ip6->sin6_len = sizeof(struct sockaddr_in6);
 959                 ip6->sin6_scope_id = listen6->sin6_scope_id;
 960                 break;
 961         default:
 962                 break;
 963         }
 964 }
 965
 966 static inline int cma_user_data_offset(enum rdma_port_space ps)
 967 {
 968         switch (ps) {
 969         case RDMA_PS_SDP:
 970                 return 0;
 971         default:
 972                 return sizeof(struct cma_hdr);
 973         }
 974 }
 975
 976 static void cma_cancel_route(struct rdma_id_private *id_priv)
 977 {
 978         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 979         case IB_LINK_LAYER_INFINIBAND:
 980                 if (id_priv->query)
 981                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 982                 break;
 983         default:
 984                 break;
 985         }
 986 }
 987
 988 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 989 {
 990         struct rdma_id_private *dev_id_priv;
 991
 992         /*
 993          * Remove from listen_any_list to prevent added devices from spawning
 994          * additional listen requests.
 995          */
 996         mutex_lock(&lock);
 997         list_del(&id_priv->list);
 998
 999         while (!list_empty(&id_priv->listen_list)) {
1000                 dev_id_priv = list_entry(id_priv->listen_list.next,
1001                                          struct rdma_id_private, listen_list);
1002                 /* sync with device removal to avoid duplicate destruction */
1003                 list_del_init(&dev_id_priv->list);
1004                 list_del(&dev_id_priv->listen_list);
1005                 mutex_unlock(&lock);
1006
1007                 rdma_destroy_id(&dev_id_priv->id);
1008                 mutex_lock(&lock);
1009         }
1010         mutex_unlock(&lock);
1011 }
1012
1013 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1014                                  enum rdma_cm_state state)
1015 {
1016         switch (state) {
1017         case RDMA_CM_ADDR_QUERY:
1018                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1019                 break;
1020         case RDMA_CM_ROUTE_QUERY:
1021                 cma_cancel_route(id_priv);
1022                 break;
1023         case RDMA_CM_LISTEN:
1024                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
1025                                 && !id_priv->cma_dev)
1026                         cma_cancel_listens(id_priv);
1027                 break;
1028         default:
1029                 break;
1030         }
1031 }
1032
1033 static void cma_release_port(struct rdma_id_private *id_priv)
1034 {
1035         struct rdma_bind_list *bind_list;
1036
1037         mutex_lock(&lock);
1038         bind_list = id_priv->bind_list;
1039         if (!bind_list) {
1040                 mutex_unlock(&lock);
1041                 return;
1042         }
1043         hlist_del(&id_priv->node);
1044         id_priv->bind_list = NULL;
1045         if (hlist_empty(&bind_list->owners)) {
1046                 idr_remove(bind_list->ps, bind_list->port);
1047                 kfree(bind_list);
1048         }
1049         mutex_unlock(&lock);
1050         if (id_priv->sock)
1051                 sock_release(id_priv->sock);
1052 }
1053
1054 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1055 {
1056         struct cma_multicast *mc;
1057
1058         while (!list_empty(&id_priv->mc_list)) {
1059                 mc = container_of(id_priv->mc_list.next,
1060                                   struct cma_multicast, list);
1061                 list_del(&mc->list);
1062                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
1063                 case IB_LINK_LAYER_INFINIBAND:
1064                         ib_sa_free_multicast(mc->multicast.ib);
1065                         kfree(mc);
1066                         break;
1067                 case IB_LINK_LAYER_ETHERNET:
1068                         kref_put(&mc->mcref, release_mc);
1069                         break;
1070                 default:
1071                         break;
1072                 }
1073         }
1074 }
1075 static void __rdma_free(struct work_struct *work)
1076 {
1077         struct rdma_id_private *id_priv;
1078         id_priv = container_of(work, struct rdma_id_private, work);
1079
1080         wait_for_completion(&id_priv->comp);
1081
1082         if (id_priv->internal_id)
1083                 cma_deref_id(id_priv->id.context);
1084
1085         kfree(id_priv->id.route.path_rec);
1086         kfree(id_priv);
1087 }
1088
1089 void rdma_destroy_id(struct rdma_cm_id *id)
1090 {
1091         struct rdma_id_private *id_priv;
1092         enum rdma_cm_state state;
1093         unsigned long flags;
1094         struct ib_cm_id *ib;
1095
1096         id_priv = container_of(id, struct rdma_id_private, id);
1097         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1098         cma_cancel_operation(id_priv, state);
1099
1100         /*
1101          * Wait for any active callback to finish.  New callbacks will find
1102          * the id_priv state set to destroying and abort.
1103          */
1104         mutex_lock(&id_priv->handler_mutex);
1105         mutex_unlock(&id_priv->handler_mutex);
1106
1107         if (id_priv->cma_dev) {
1108                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1109                 case RDMA_TRANSPORT_IB:
1110                         spin_lock_irqsave(&id_priv->cm_lock, flags);
1111                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib)) {
1112                                 ib = id_priv->cm_id.ib;
1113                                 id_priv->cm_id.ib = NULL;
1114                                 spin_unlock_irqrestore(&id_priv->cm_lock, flags);
1115                                 ib_destroy_cm_id(ib);
1116                         } else
1117                                 spin_unlock_irqrestore(&id_priv->cm_lock, flags);
1118                         break;
1119                 case RDMA_TRANSPORT_IWARP:
1120                 case RDMA_TRANSPORT_SCIF:
1121                         if (id_priv->cm_id.iw)
1122                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1123                         break;
1124                 default:
1125                         break;
1126                 }
1127                 cma_leave_mc_groups(id_priv);
1128                 cma_release_dev(id_priv);
1129         }
1130
1131         cma_release_port(id_priv);
1132         cma_deref_id(id_priv);
1133         INIT_WORK(&id_priv->work, __rdma_free);
1134         queue_work(cma_free_wq, &id_priv->work);
1135 }
1136 EXPORT_SYMBOL(rdma_destroy_id);
1137
1138 static int cma_rep_recv(struct rdma_id_private *id_priv)
1139 {
1140         int ret;
1141
1142         ret = cma_modify_qp_rtr(id_priv, NULL);
1143         if (ret)
1144                 goto reject;
1145
1146         ret = cma_modify_qp_rts(id_priv, NULL);
1147         if (ret)
1148                 goto reject;
1149
1150         cma_dbg(id_priv, "sending RTU\n");
1151         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1152         if (ret)
1153                 goto reject;
1154
1155         return 0;
1156 reject:
1157         cma_modify_qp_err(id_priv);
1158         cma_dbg(id_priv, "sending REJ\n");
1159         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1160                        NULL, 0, NULL, 0);
1161         return ret;
1162 }
1163
1164 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
1165 {
1166         if (id_priv->id.ps == RDMA_PS_SDP &&
1167             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
1168             SDP_MAJ_VERSION)
1169                 return -EINVAL;
1170
1171         return 0;
1172 }
1173
1174 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1175                                    struct ib_cm_rep_event_param *rep_data,
1176                                    void *private_data)
1177 {
1178         event->param.conn.private_data = private_data;
1179         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1180         event->param.conn.responder_resources = rep_data->responder_resources;
1181         event->param.conn.initiator_depth = rep_data->initiator_depth;
1182         event->param.conn.flow_control = rep_data->flow_control;
1183         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1184         event->param.conn.srq = rep_data->srq;
1185         event->param.conn.qp_num = rep_data->remote_qpn;
1186 }
1187
1188 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1189 {
1190         struct rdma_id_private *id_priv = cm_id->context;
1191         struct rdma_cm_event event;
1192         int ret = 0;
1193
1194         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1195                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1196             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1197                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1198                 return 0;
1199         memset(&event, 0, sizeof event);
1200         switch (ib_event->event) {
1201         case IB_CM_REQ_ERROR:
1202         case IB_CM_REP_ERROR:
1203                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1204                 event.status = -ETIMEDOUT;
1205                 break;
1206         case IB_CM_REP_RECEIVED:
1207                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1208                 if (event.status)
1209                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1210                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1211                         event.status = cma_rep_recv(id_priv);
1212                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1213                                                      RDMA_CM_EVENT_ESTABLISHED;
1214                 } else
1215                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1216                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1217                                        ib_event->private_data);
1218                 break;
1219         case IB_CM_RTU_RECEIVED:
1220         case IB_CM_USER_ESTABLISHED:
1221                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1222                 break;
1223         case IB_CM_DREQ_ERROR:
1224                 event.status = -ETIMEDOUT; /* fall through */
1225         case IB_CM_DREQ_RECEIVED:
1226         case IB_CM_DREP_RECEIVED:
1227                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1228                                    RDMA_CM_DISCONNECT))
1229                         goto out;
1230                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1231                 break;
1232         case IB_CM_TIMEWAIT_EXIT:
1233                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1234                 break;
1235         case IB_CM_MRA_RECEIVED:
1236                 /* ignore event */
1237                 goto out;
1238         case IB_CM_REJ_RECEIVED:
1239                 cma_modify_qp_err(id_priv);
1240                 event.status = ib_event->param.rej_rcvd.reason;
1241                 event.event = RDMA_CM_EVENT_REJECTED;
1242                 event.param.conn.private_data = ib_event->private_data;
1243                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1244                 break;
1245         default:
1246                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1247                        ib_event->event);
1248                 goto out;
1249         }
1250
1251         ret = id_priv->id.event_handler(&id_priv->id, &event);
1252         if (ret) {
1253                 /* Destroy the CM ID by returning a non-zero value. */
1254                 id_priv->cm_id.ib = NULL;
1255                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1256                 mutex_unlock(&id_priv->handler_mutex);
1257                 rdma_destroy_id(&id_priv->id);
1258                 return ret;
1259         }
1260 out:
1261         mutex_unlock(&id_priv->handler_mutex);
1262         return ret;
1263 }
1264
1265 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1266                                                struct ib_cm_event *ib_event)
1267 {
1268         struct rdma_id_private *id_priv;
1269         struct rdma_cm_id *id;
1270         struct rdma_route *rt;
1271         union cma_ip_addr *src, *dst;
1272         __be16 port;
1273         u8 ip_ver;
1274         int ret;
1275
1276         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1277                              &ip_ver, &port, &src, &dst))
1278                 return NULL;
1279
1280         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1281                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1282         if (IS_ERR(id))
1283                 return NULL;
1284
1285         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1286                           ip_ver, port, src, dst);
1287
1288         rt = &id->route;
1289         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1290         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1291                                GFP_KERNEL);
1292         if (!rt->path_rec)
1293                 goto err;
1294
1295         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1296         if (rt->num_paths == 2)
1297                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1298
1299         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1300                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1301                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1302                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1303         } else {
1304                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1305                                         &rt->addr.dev_addr, NULL);
1306                 if (ret)
1307                         goto err;
1308         }
1309         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1310
1311         id_priv = container_of(id, struct rdma_id_private, id);
1312         id_priv->state = RDMA_CM_CONNECT;
1313         return id_priv;
1314
1315 err:
1316         rdma_destroy_id(id);
1317         return NULL;
1318 }
1319
1320 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1321                                               struct ib_cm_event *ib_event)
1322 {
1323         struct rdma_id_private *id_priv;
1324         struct rdma_cm_id *id;
1325         union cma_ip_addr *src, *dst;
1326         __be16 port;
1327         u8 ip_ver;
1328         int ret;
1329
1330         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1331                             listen_id->ps, IB_QPT_UD);
1332         if (IS_ERR(id))
1333                 return NULL;
1334
1335
1336         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1337                              &ip_ver, &port, &src, &dst))
1338                 goto err;
1339
1340         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1341                           ip_ver, port, src, dst);
1342
1343         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1344                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1345                                         &id->route.addr.dev_addr, NULL);
1346                 if (ret)
1347                         goto err;
1348         }
1349
1350         id_priv = container_of(id, struct rdma_id_private, id);
1351         id_priv->state = RDMA_CM_CONNECT;
1352         return id_priv;
1353 err:
1354         rdma_destroy_id(id);
1355         return NULL;
1356 }
1357
1358 static void cma_set_req_event_data(struct rdma_cm_event *event,
1359                                    struct ib_cm_req_event_param *req_data,
1360                                    void *private_data, int offset)
1361 {
1362         event->param.conn.private_data = private_data + offset;
1363         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1364         event->param.conn.responder_resources = req_data->responder_resources;
1365         event->param.conn.initiator_depth = req_data->initiator_depth;
1366         event->param.conn.flow_control = req_data->flow_control;
1367         event->param.conn.retry_count = req_data->retry_count;
1368         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1369         event->param.conn.srq = req_data->srq;
1370         event->param.conn.qp_num = req_data->remote_qpn;
1371 }
1372
1373 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1374 {
1375         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1376                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1377                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1378                  (id->qp_type == IB_QPT_UD)) ||
1379                 (!id->qp_type));
1380 }
1381
1382 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1383 {
1384         struct rdma_id_private *listen_id, *conn_id;
1385         struct rdma_cm_event event;
1386         int offset, ret;
1387         u8 smac[ETH_ALEN];
1388         u8 alt_smac[ETH_ALEN];
1389         u8 *psmac = smac;
1390         u8 *palt_smac = alt_smac;
1391         int is_iboe = ((rdma_node_get_transport(cm_id->device->node_type) ==
1392                         RDMA_TRANSPORT_IB) &&
1393                        (rdma_port_get_link_layer(cm_id->device,
1394                         ib_event->param.req_rcvd.port) ==
1395                         IB_LINK_LAYER_ETHERNET));
1396         int is_sidr = 0;
1397
1398         listen_id = cm_id->context;
1399         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1400                 return -EINVAL;
1401
1402         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1403                 return -ECONNABORTED;
1404
1405         memset(&event, 0, sizeof event);
1406         offset = cma_user_data_offset(listen_id->id.ps);
1407         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1408         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1409                 is_sidr = 1;
1410                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1411                 event.param.ud.private_data = ib_event->private_data + offset;
1412                 event.param.ud.private_data_len =
1413                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1414         } else {
1415                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1416                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1417                                        ib_event->private_data, offset);
1418         }
1419         if (!conn_id) {
1420                 ret = -ENOMEM;
1421                 goto err1;
1422         }
1423
1424         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1425         ret = cma_acquire_dev(conn_id);
1426         if (ret)
1427                 goto err2;
1428
1429         conn_id->cm_id.ib = cm_id;
1430         cm_id->context = conn_id;
1431         cm_id->cm_handler = cma_ib_handler;
1432
1433         /*
1434          * Protect against the user destroying conn_id from another thread
1435          * until we're done accessing it.
1436          */
1437         atomic_inc(&conn_id->refcount);
1438         ret = conn_id->id.event_handler(&conn_id->id, &event);
1439         if (ret)
1440                 goto err3;
1441
1442         if (is_iboe && !is_sidr) {
1443                 u32 scope_id = rdma_get_ipv6_scope_id(cm_id->device,
1444                     ib_event->param.req_rcvd.port);
1445
1446                 if (ib_event->param.req_rcvd.primary_path != NULL)
1447                         rdma_addr_find_smac_by_sgid(
1448                                 &ib_event->param.req_rcvd.primary_path->sgid,
1449                                 psmac, NULL, scope_id);
1450                 else
1451                         psmac = NULL;
1452                 if (ib_event->param.req_rcvd.alternate_path != NULL)
1453                         rdma_addr_find_smac_by_sgid(
1454                                 &ib_event->param.req_rcvd.alternate_path->sgid,
1455                                 palt_smac, NULL, scope_id);
1456                 else
1457                         palt_smac = NULL;
1458         }
1459                 /*
1460                  * Acquire mutex to prevent user executing rdma_destroy_id()
1461                  * while we're accessing the cm_id.
1462                  */
1463                 mutex_lock(&lock);
1464         if (is_iboe && !is_sidr)
1465                 ib_update_cm_av(cm_id, psmac, palt_smac);
1466         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD)) {
1467                 cma_dbg(container_of(&conn_id->id, struct rdma_id_private, id), "sending MRA\n");
1468                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1469         }
1470                 mutex_unlock(&lock);
1471                 mutex_unlock(&conn_id->handler_mutex);
1472         mutex_unlock(&listen_id->handler_mutex);
1473         cma_deref_id(conn_id);
1474         return 0;
1475
1476 err3:
1477         cma_deref_id(conn_id);
1478         /* Destroy the CM ID by returning a non-zero value. */
1479         conn_id->cm_id.ib = NULL;
1480 err2:
1481         cma_exch(conn_id, RDMA_CM_DESTROYING);
1482         mutex_unlock(&conn_id->handler_mutex);
1483 err1:
1484         mutex_unlock(&listen_id->handler_mutex);
1485         if (conn_id)
1486                 rdma_destroy_id(&conn_id->id);
1487         return ret;
1488 }
1489
1490 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1491 {
1492         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1493 }
1494
1495 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1496                                  struct ib_cm_compare_data *compare)
1497 {
1498         struct cma_hdr *cma_data, *cma_mask;
1499         struct sdp_hh *sdp_data, *sdp_mask;
1500         __be32 ip4_addr;
1501         struct in6_addr ip6_addr;
1502
1503         memset(compare, 0, sizeof *compare);
1504         cma_data = (void *) compare->data;
1505         cma_mask = (void *) compare->mask;
1506         sdp_data = (void *) compare->data;
1507         sdp_mask = (void *) compare->mask;
1508
1509         switch (addr->sa_family) {
1510         case AF_INET:
1511                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1512                 if (ps == RDMA_PS_SDP) {
1513                         sdp_set_ip_ver(sdp_data, 4);
1514                         sdp_set_ip_ver(sdp_mask, 0xF);
1515                         if (!cma_any_addr(addr)) {
1516                                 sdp_data->dst_addr.ip4.addr = ip4_addr;
1517                                 sdp_mask->dst_addr.ip4.addr = htonl(~0);
1518                         }
1519                 } else {
1520                         cma_set_ip_ver(cma_data, 4);
1521                         cma_set_ip_ver(cma_mask, 0xF);
1522                         if (!cma_any_addr(addr)) {
1523                                 cma_data->dst_addr.ip4.addr = ip4_addr;
1524                                 cma_mask->dst_addr.ip4.addr = htonl(~0);
1525                         }
1526                 }
1527                 break;
1528         case AF_INET6:
1529                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1530                 if (ps == RDMA_PS_SDP) {
1531                         sdp_set_ip_ver(sdp_data, 6);
1532                         sdp_set_ip_ver(sdp_mask, 0xF);
1533                         if (!cma_any_addr(addr)) {
1534                                 sdp_data->dst_addr.ip6 = ip6_addr;
1535                                 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1536                                        sizeof(sdp_mask->dst_addr.ip6));
1537                         }
1538                 } else {
1539                         cma_set_ip_ver(cma_data, 6);
1540                         cma_set_ip_ver(cma_mask, 0xF);
1541                         if (!cma_any_addr(addr)) {
1542                                 cma_data->dst_addr.ip6 = ip6_addr;
1543                                 memset(&cma_mask->dst_addr.ip6, 0xFF,
1544                                        sizeof(cma_mask->dst_addr.ip6));
1545                         }
1546                 }
1547                 break;
1548         default:
1549                 break;
1550         }
1551 }
1552
1553 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1554 {
1555         struct rdma_id_private *id_priv = iw_id->context;
1556         struct rdma_cm_event event;
1557         struct sockaddr_in *sin;
1558         int ret = 0;
1559
1560         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1561                 return 0;
1562
1563         memset(&event, 0, sizeof event);
1564         switch (iw_event->event) {
1565         case IW_CM_EVENT_CLOSE:
1566                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1567                 break;
1568         case IW_CM_EVENT_CONNECT_REPLY:
1569                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1570                 *sin = iw_event->local_addr;
1571                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1572                 *sin = iw_event->remote_addr;
1573                 switch ((int)iw_event->status) {
1574                 case 0:
1575                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1576                         event.param.conn.initiator_depth = iw_event->ird;
1577                         event.param.conn.responder_resources = iw_event->ord;
1578                         break;
1579                 case -ECONNRESET:
1580                 case -ECONNREFUSED:
1581                         event.event = RDMA_CM_EVENT_REJECTED;
1582                         break;
1583                 case -ETIMEDOUT:
1584                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1585                         break;
1586                 default:
1587                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1588                         break;
1589                 }
1590                 break;
1591         case IW_CM_EVENT_ESTABLISHED:
1592                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1593                 event.param.conn.initiator_depth = iw_event->ird;
1594                 event.param.conn.responder_resources = iw_event->ord;
1595                 break;
1596         default:
1597                 BUG_ON(1);
1598         }
1599
1600         event.status = iw_event->status;
1601         event.param.conn.private_data = iw_event->private_data;
1602         event.param.conn.private_data_len = iw_event->private_data_len;
1603         ret = id_priv->id.event_handler(&id_priv->id, &event);
1604         if (ret) {
1605                 /* Destroy the CM ID by returning a non-zero value. */
1606                 id_priv->cm_id.iw = NULL;
1607                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1608                 mutex_unlock(&id_priv->handler_mutex);
1609                 rdma_destroy_id(&id_priv->id);
1610                 return ret;
1611         }
1612
1613         mutex_unlock(&id_priv->handler_mutex);
1614         return ret;
1615 }
1616
1617 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1618                                struct iw_cm_event *iw_event)
1619 {
1620         struct rdma_cm_id *new_cm_id;
1621         struct rdma_id_private *listen_id, *conn_id;
1622         struct sockaddr_in *sin;
1623         struct net_device *dev = NULL;
1624         struct rdma_cm_event event;
1625         int ret;
1626         struct ib_device_attr attr;
1627
1628         listen_id = cm_id->context;
1629         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1630                 return -ECONNABORTED;
1631
1632         /* Create a new RDMA id for the new IW CM ID */
1633         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1634                                    listen_id->id.context,
1635                                    RDMA_PS_TCP, IB_QPT_RC);
1636         if (IS_ERR(new_cm_id)) {
1637                 ret = -ENOMEM;
1638                 goto out;
1639         }
1640         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1641         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1642         conn_id->state = RDMA_CM_CONNECT;
1643
1644         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1645         if (!dev) {
1646                 ret = -EADDRNOTAVAIL;
1647                 mutex_unlock(&conn_id->handler_mutex);
1648                 rdma_destroy_id(new_cm_id);
1649                 goto out;
1650         }
1651         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1652         if (ret) {
1653                 mutex_unlock(&conn_id->handler_mutex);
1654                 rdma_destroy_id(new_cm_id);
1655                 goto out;
1656         }
1657
1658         ret = cma_acquire_dev(conn_id);
1659         if (ret) {
1660                 mutex_unlock(&conn_id->handler_mutex);
1661                 rdma_destroy_id(new_cm_id);
1662                 goto out;
1663         }
1664
1665         conn_id->cm_id.iw = cm_id;
1666         cm_id->context = conn_id;
1667         cm_id->cm_handler = cma_iw_handler;
1668
1669         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1670         *sin = iw_event->local_addr;
1671         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1672         *sin = iw_event->remote_addr;
1673
1674         ret = ib_query_device(conn_id->id.device, &attr);
1675         if (ret) {
1676                 mutex_unlock(&conn_id->handler_mutex);
1677                 rdma_destroy_id(new_cm_id);
1678                 goto out;
1679         }
1680
1681         memset(&event, 0, sizeof event);
1682         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1683         event.param.conn.private_data = iw_event->private_data;
1684         event.param.conn.private_data_len = iw_event->private_data_len;
1685         event.param.conn.initiator_depth = iw_event->ird;
1686         event.param.conn.responder_resources = iw_event->ord;
1687
1688         /*
1689          * Protect against the user destroying conn_id from another thread
1690          * until we're done accessing it.
1691          */
1692         atomic_inc(&conn_id->refcount);
1693         ret = conn_id->id.event_handler(&conn_id->id, &event);
1694         if (ret) {
1695                 /* User wants to destroy the CM ID */
1696                 conn_id->cm_id.iw = NULL;
1697                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1698                 mutex_unlock(&conn_id->handler_mutex);
1699                 cma_deref_id(conn_id);
1700                 rdma_destroy_id(&conn_id->id);
1701                 goto out;
1702         }
1703
1704         mutex_unlock(&conn_id->handler_mutex);
1705         cma_deref_id(conn_id);
1706
1707 out:
1708         if (dev)
1709                 dev_put(dev);
1710         mutex_unlock(&listen_id->handler_mutex);
1711         return ret;
1712 }
1713
1714 static int cma_ib_listen(struct rdma_id_private *id_priv)
1715 {
1716         struct ib_cm_compare_data compare_data;
1717         struct sockaddr *addr;
1718         struct ib_cm_id *id;
1719         __be64 svc_id;
1720         int ret;
1721
1722         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1723         if (IS_ERR(id))
1724                 return PTR_ERR(id);
1725
1726         id_priv->cm_id.ib = id;
1727
1728         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1729         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1730         if (cma_any_addr(addr) && !id_priv->afonly)
1731                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1732         else {
1733                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1734                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1735         }
1736
1737         if (ret) {
1738                 ib_destroy_cm_id(id_priv->cm_id.ib);
1739                 id_priv->cm_id.ib = NULL;
1740         }
1741
1742         return ret;
1743 }
1744
1745 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1746 {
1747         int ret;
1748         struct sockaddr_in *sin;
1749         struct iw_cm_id *id;
1750
1751         id = iw_create_cm_id(id_priv->id.device,
1752                                             id_priv->sock,
1753                                             iw_conn_req_handler,
1754                                             id_priv);
1755         if (IS_ERR(id))
1756                 return PTR_ERR(id);
1757
1758         id_priv->cm_id.iw = id;
1759
1760         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1761         id_priv->cm_id.iw->local_addr = *sin;
1762
1763         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1764
1765         if (ret) {
1766                 iw_destroy_cm_id(id_priv->cm_id.iw);
1767                 id_priv->cm_id.iw = NULL;
1768         }
1769
1770         return ret;
1771 }
1772
1773 static int cma_listen_handler(struct rdma_cm_id *id,
1774                               struct rdma_cm_event *event)
1775 {
1776         struct rdma_id_private *id_priv = id->context;
1777
1778         id->context = id_priv->id.context;
1779         id->event_handler = id_priv->id.event_handler;
1780         return id_priv->id.event_handler(id, event);
1781 }
1782
1783 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1784                               struct cma_device *cma_dev)
1785 {
1786         struct rdma_id_private *dev_id_priv;
1787         struct rdma_cm_id *id;
1788         int ret;
1789
1790         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1791                             id_priv->id.qp_type);
1792         if (IS_ERR(id))
1793                 return;
1794
1795         dev_id_priv = container_of(id, struct rdma_id_private, id);
1796
1797         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1798         dev_id_priv->sock = id_priv->sock;
1799         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1800                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1801
1802         cma_attach_to_dev(dev_id_priv, cma_dev);
1803         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1804         atomic_inc(&id_priv->refcount);
1805         dev_id_priv->internal_id = 1;
1806         dev_id_priv->afonly = id_priv->afonly;
1807
1808         ret = rdma_listen(id, id_priv->backlog);
1809         if (ret)
1810                 cma_warn(id_priv, "cma_listen_on_dev, error %d, listening on device %s\n", ret, cma_dev->device->name);
1811 }
1812
1813 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1814 {
1815         struct cma_device *cma_dev;
1816
1817         mutex_lock(&lock);
1818         list_add_tail(&id_priv->list, &listen_any_list);
1819         list_for_each_entry(cma_dev, &dev_list, list)
1820                 cma_listen_on_dev(id_priv, cma_dev);
1821         mutex_unlock(&lock);
1822 }
1823
1824 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1825 {
1826         struct rdma_id_private *id_priv;
1827
1828         id_priv = container_of(id, struct rdma_id_private, id);
1829         id_priv->tos = (u8) tos;
1830 }
1831 EXPORT_SYMBOL(rdma_set_service_type);
1832
1833 void rdma_set_timeout(struct rdma_cm_id *id, int timeout)
1834 {
1835         struct rdma_id_private *id_priv;
1836
1837         id_priv = container_of(id, struct rdma_id_private, id);
1838         id_priv->qp_timeout = (u8) timeout;
1839 }
1840 EXPORT_SYMBOL(rdma_set_timeout);
1841
1842 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1843                               void *context)
1844 {
1845         struct cma_work *work = context;
1846         struct rdma_route *route;
1847
1848         route = &work->id->id.route;
1849
1850         if (!status) {
1851                 route->num_paths = 1;
1852                 *route->path_rec = *path_rec;
1853         } else {
1854                 work->old_state = RDMA_CM_ROUTE_QUERY;
1855                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1856                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1857                 work->event.status = status;
1858         }
1859
1860         queue_work(cma_wq, &work->work);
1861 }
1862
1863 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1864                               struct cma_work *work)
1865 {
1866         struct rdma_addr *addr = &id_priv->id.route.addr;
1867         struct ib_sa_path_rec path_rec;
1868         ib_sa_comp_mask comp_mask;
1869         struct sockaddr_in6 *sin6;
1870
1871         memset(&path_rec, 0, sizeof path_rec);
1872         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1873         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1874         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1875         path_rec.numb_path = 1;
1876         path_rec.reversible = 1;
1877         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1878                                                         (struct sockaddr *) &addr->dst_addr);
1879
1880         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1881                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1882                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1883
1884         if (addr->src_addr.ss_family == AF_INET) {
1885                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1886                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1887         } else {
1888                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1889                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1890                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1891         }
1892
1893         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1894                                                id_priv->id.port_num, &path_rec,
1895                                                comp_mask, timeout_ms,
1896                                                GFP_KERNEL, cma_query_handler,
1897                                                work, &id_priv->query);
1898
1899         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1900 }
1901
1902 static void cma_work_handler(struct work_struct *_work)
1903 {
1904         struct cma_work *work = container_of(_work, struct cma_work, work);
1905         struct rdma_id_private *id_priv = work->id;
1906         int destroy = 0;
1907
1908         mutex_lock(&id_priv->handler_mutex);
1909         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1910                 goto out;
1911
1912         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1913                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1914                 destroy = 1;
1915         }
1916 out:
1917         mutex_unlock(&id_priv->handler_mutex);
1918         cma_deref_id(id_priv);
1919         if (destroy)
1920                 rdma_destroy_id(&id_priv->id);
1921         kfree(work);
1922 }
1923
1924 static void cma_ndev_work_handler(struct work_struct *_work)
1925 {
1926         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1927         struct rdma_id_private *id_priv = work->id;
1928         int destroy = 0;
1929
1930         mutex_lock(&id_priv->handler_mutex);
1931         if (id_priv->state == RDMA_CM_DESTROYING ||
1932             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1933                 goto out;
1934
1935         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1936                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1937                 destroy = 1;
1938         }
1939
1940 out:
1941         mutex_unlock(&id_priv->handler_mutex);
1942         cma_deref_id(id_priv);
1943         if (destroy)
1944                 rdma_destroy_id(&id_priv->id);
1945         kfree(work);
1946 }
1947
1948 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1949 {
1950         struct rdma_route *route = &id_priv->id.route;
1951         struct cma_work *work;
1952         int ret;
1953
1954         work = kzalloc(sizeof *work, GFP_KERNEL);
1955         if (!work)
1956                 return -ENOMEM;
1957
1958         work->id = id_priv;
1959         INIT_WORK(&work->work, cma_work_handler);
1960         work->old_state = RDMA_CM_ROUTE_QUERY;
1961         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1962         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1963
1964         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1965         if (!route->path_rec) {
1966                 ret = -ENOMEM;
1967                 goto err1;
1968         }
1969
1970         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1971         if (ret)
1972                 goto err2;
1973
1974         return 0;
1975 err2:
1976         kfree(route->path_rec);
1977         route->path_rec = NULL;
1978 err1:
1979         kfree(work);
1980         return ret;
1981 }
1982
1983 int rdma_set_ib_paths(struct rdma_cm_id *id,
1984                       struct ib_sa_path_rec *path_rec, int num_paths)
1985 {
1986         struct rdma_id_private *id_priv;
1987         int ret;
1988
1989         id_priv = container_of(id, struct rdma_id_private, id);
1990         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1991                            RDMA_CM_ROUTE_RESOLVED))
1992                 return -EINVAL;
1993
1994         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1995                                      GFP_KERNEL);
1996         if (!id->route.path_rec) {
1997                 ret = -ENOMEM;
1998                 goto err;
1999         }
2000
2001         id->route.num_paths = num_paths;
2002         return 0;
2003 err:
2004         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2005         return ret;
2006 }
2007 EXPORT_SYMBOL(rdma_set_ib_paths);
2008
2009 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2010 {
2011         struct cma_work *work;
2012
2013         work = kzalloc(sizeof *work, GFP_KERNEL);
2014         if (!work)
2015                 return -ENOMEM;
2016
2017         work->id = id_priv;
2018         INIT_WORK(&work->work, cma_work_handler);
2019         work->old_state = RDMA_CM_ROUTE_QUERY;
2020         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2021         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2022         queue_work(cma_wq, &work->work);
2023         return 0;
2024 }
2025
2026 static u8 tos_to_sl(u8 tos)
2027 {
2028         return def_prec2sl & 7;
2029 }
2030
2031 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2032 {
2033         struct rdma_route *route = &id_priv->id.route;
2034         struct rdma_addr *addr = &route->addr;
2035         struct cma_work *work;
2036         int ret;
2037         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
2038         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
2039         struct net_device *ndev = NULL;
2040
2041
2042         if (src_addr->sin_family != dst_addr->sin_family)
2043                 return -EINVAL;
2044
2045         work = kzalloc(sizeof *work, GFP_KERNEL);
2046         if (!work)
2047                 return -ENOMEM;
2048
2049         work->id = id_priv;
2050         INIT_WORK(&work->work, cma_work_handler);
2051
2052         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2053         if (!route->path_rec) {
2054                 ret = -ENOMEM;
2055                 goto err1;
2056         }
2057
2058         route->num_paths = 1;
2059
2060         if (addr->dev_addr.bound_dev_if)
2061                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2062         if (!ndev) {
2063                 ret = -ENODEV;
2064                 goto err2;
2065         }
2066
2067         route->path_rec->vlan_id = rdma_vlan_dev_vlan_id(ndev);
2068         memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2069         memcpy(route->path_rec->smac, IF_LLADDR(ndev), ndev->if_addrlen);
2070
2071
2072         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2073                     &route->path_rec->sgid);
2074         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2075                     &route->path_rec->dgid);
2076
2077         route->path_rec->hop_limit = 1;
2078         route->path_rec->reversible = 1;
2079         route->path_rec->pkey = cpu_to_be16(0xffff);
2080         route->path_rec->mtu_selector = IB_SA_EQ;
2081         route->path_rec->sl = tos_to_sl(id_priv->tos);
2082
2083         route->path_rec->mtu = iboe_get_mtu(ndev->if_mtu);
2084         route->path_rec->rate_selector = IB_SA_EQ;
2085         route->path_rec->rate = iboe_get_rate(ndev);
2086         dev_put(ndev);
2087         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2088         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2089         if (!route->path_rec->mtu) {
2090                 ret = -EINVAL;
2091                 goto err2;
2092         }
2093
2094         work->old_state = RDMA_CM_ROUTE_QUERY;
2095         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2096         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2097         work->event.status = 0;
2098
2099         queue_work(cma_wq, &work->work);
2100
2101         return 0;
2102
2103 err2:
2104         kfree(route->path_rec);
2105         route->path_rec = NULL;
2106 err1:
2107         kfree(work);
2108         return ret;
2109 }
2110
2111 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2112 {
2113         struct rdma_id_private *id_priv;
2114         int ret;
2115
2116         id_priv = container_of(id, struct rdma_id_private, id);
2117         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2118                 return -EINVAL;
2119
2120         atomic_inc(&id_priv->refcount);
2121         switch (rdma_node_get_transport(id->device->node_type)) {
2122         case RDMA_TRANSPORT_IB:
2123                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
2124                 case IB_LINK_LAYER_INFINIBAND:
2125                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
2126                         break;
2127                 case IB_LINK_LAYER_ETHERNET:
2128                         ret = cma_resolve_iboe_route(id_priv);
2129                         break;
2130                 default:
2131                         ret = -ENOSYS;
2132                 }
2133                 break;
2134         case RDMA_TRANSPORT_IWARP:
2135         case RDMA_TRANSPORT_SCIF:
2136                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2137                 break;
2138         default:
2139                 ret = -ENOSYS;
2140                 break;
2141         }
2142         if (ret)
2143                 goto err;
2144
2145         return 0;
2146 err:
2147         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2148         cma_deref_id(id_priv);
2149         return ret;
2150 }
2151 EXPORT_SYMBOL(rdma_resolve_route);
2152
2153 int rdma_enable_apm(struct rdma_cm_id *id, enum alt_path_type alt_type)
2154 {
2155         /* APM is not supported yet */
2156         return -EINVAL;
2157 }
2158 EXPORT_SYMBOL(rdma_enable_apm);
2159
2160 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2161 {
2162         struct cma_device *cma_dev;
2163         struct ib_port_attr port_attr;
2164         union ib_gid gid;
2165         u16 pkey;
2166         int ret;
2167         u8 p;
2168
2169         mutex_lock(&lock);
2170         if (list_empty(&dev_list)) {
2171                 ret = -ENODEV;
2172                 goto out;
2173         }
2174         list_for_each_entry(cma_dev, &dev_list, list)
2175                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
2176                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
2177                             port_attr.state == IB_PORT_ACTIVE)
2178                                 goto port_found;
2179
2180         p = 1;
2181         cma_dev = list_entry(dev_list.next, struct cma_device, list);
2182
2183 port_found:
2184         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2185         if (ret)
2186                 goto out;
2187
2188         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2189         if (ret)
2190                 goto out;
2191
2192         id_priv->id.route.addr.dev_addr.dev_type =
2193                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2194                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2195
2196         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2197         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2198         id_priv->id.port_num = p;
2199         cma_attach_to_dev(id_priv, cma_dev);
2200 out:
2201         mutex_unlock(&lock);
2202         return ret;
2203 }
2204
2205 static void addr_handler(int status, struct sockaddr *src_addr,
2206                          struct rdma_dev_addr *dev_addr, void *context)
2207 {
2208         struct rdma_id_private *id_priv = context;
2209         struct rdma_cm_event event;
2210
2211         memset(&event, 0, sizeof event);
2212         mutex_lock(&id_priv->handler_mutex);
2213         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2214                            RDMA_CM_ADDR_RESOLVED))
2215                 goto out;
2216
2217         memcpy(&id_priv->id.route.addr.src_addr, src_addr,
2218                ip_addr_size(src_addr));
2219         if (!status && !id_priv->cma_dev)
2220                 status = cma_acquire_dev(id_priv);
2221
2222         if (status) {
2223                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2224                                    RDMA_CM_ADDR_BOUND))
2225                         goto out;
2226                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2227                 event.status = status;
2228         } else
2229                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2230
2231         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2232                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2233                 mutex_unlock(&id_priv->handler_mutex);
2234                 cma_deref_id(id_priv);
2235                 rdma_destroy_id(&id_priv->id);
2236                 return;
2237         }
2238 out:
2239         mutex_unlock(&id_priv->handler_mutex);
2240         cma_deref_id(id_priv);
2241 }
2242
2243 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2244 {
2245         struct cma_work *work;
2246         struct sockaddr *src, *dst;
2247         union ib_gid gid;
2248         int ret;
2249
2250         work = kzalloc(sizeof *work, GFP_KERNEL);
2251         if (!work)
2252                 return -ENOMEM;
2253
2254         if (!id_priv->cma_dev) {
2255                 ret = cma_bind_loopback(id_priv);
2256                 if (ret)
2257                         goto err;
2258         }
2259
2260         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2261         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2262
2263         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2264         if (cma_zero_addr(src)) {
2265                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2266                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2267                         ((struct sockaddr_in *)src)->sin_addr =
2268                                 ((struct sockaddr_in *)dst)->sin_addr;
2269                 } else {
2270                         ((struct sockaddr_in6 *)src)->sin6_addr =
2271                                 ((struct sockaddr_in6 *)dst)->sin6_addr;
2272                 }
2273         }
2274
2275         work->id = id_priv;
2276         INIT_WORK(&work->work, cma_work_handler);
2277         work->old_state = RDMA_CM_ADDR_QUERY;
2278         work->new_state = RDMA_CM_ADDR_RESOLVED;
2279         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2280         queue_work(cma_wq, &work->work);
2281         return 0;
2282 err:
2283         kfree(work);
2284         return ret;
2285 }
2286
2287 static int cma_resolve_scif(struct rdma_id_private *id_priv)
2288 {
2289         struct cma_work *work;
2290
2291         work = kzalloc(sizeof *work, GFP_KERNEL);
2292         if (!work)
2293                 return -ENOMEM;
2294
2295         /* we probably can leave it empty here */
2296
2297         work->id = id_priv;
2298         INIT_WORK(&work->work, cma_work_handler);
2299         work->old_state = RDMA_CM_ADDR_QUERY;
2300         work->new_state = RDMA_CM_ADDR_RESOLVED;
2301         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2302         queue_work(cma_wq, &work->work);
2303         return 0;
2304 }
2305
2306 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2307                          struct sockaddr *dst_addr)
2308 {
2309         if (!src_addr || !src_addr->sa_family) {
2310                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2311                 src_addr->sa_family = dst_addr->sa_family;
2312 #ifdef INET6
2313                 if (dst_addr->sa_family == AF_INET6) {
2314                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2315                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2316                 }
2317 #endif
2318         }
2319         if (!cma_any_addr(src_addr))
2320                 return rdma_bind_addr(id, src_addr);
2321         else {
2322 #if defined(INET6) || defined(INET)
2323                 union {
2324 #ifdef INET
2325                         struct sockaddr_in in;
2326 #endif
2327 #ifdef INET6
2328                         struct sockaddr_in6 in6;
2329 #endif
2330                 } addr;
2331 #endif
2332
2333                 switch(dst_addr->sa_family) {
2334 #ifdef INET
2335                 case AF_INET:
2336                         memset(&addr.in, 0, sizeof(addr.in));
2337                         addr.in.sin_family = dst_addr->sa_family;
2338                         addr.in.sin_len = sizeof(addr.in);
2339                         return rdma_bind_addr(id, (struct sockaddr *)&addr.in);
2340 #endif
2341 #ifdef INET6
2342                 case AF_INET6:
2343                         memset(&addr.in6, 0, sizeof(addr.in6));
2344                         addr.in6.sin6_family = dst_addr->sa_family;
2345                         addr.in6.sin6_len = sizeof(addr.in6);
2346                         addr.in6.sin6_scope_id =
2347                             ((struct sockaddr_in6 *)dst_addr)->sin6_scope_id;
2348                         return rdma_bind_addr(id, (struct sockaddr *)&addr.in6);
2349 #endif
2350                 default:
2351                         return -EINVAL;
2352                 }
2353         }
2354 }
2355
2356 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2357                       struct sockaddr *dst_addr, int timeout_ms)
2358 {
2359         struct rdma_id_private *id_priv;
2360         int ret;
2361
2362         id_priv = container_of(id, struct rdma_id_private, id);
2363         if (id_priv->state == RDMA_CM_IDLE) {
2364                 ret = cma_bind_addr(id, src_addr, dst_addr);
2365                 if (ret)
2366                         return ret;
2367         }
2368
2369         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2370                 return -EINVAL;
2371
2372         atomic_inc(&id_priv->refcount);
2373         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2374         if (cma_any_addr(dst_addr))
2375                 ret = cma_resolve_loopback(id_priv);
2376         else if (id_priv->id.device &&
2377                 rdma_node_get_transport(id_priv->id.device->node_type) == RDMA_TRANSPORT_SCIF)
2378                 ret = cma_resolve_scif(id_priv);
2379         else
2380                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2381                                       dst_addr, &id->route.addr.dev_addr,
2382                                       timeout_ms, addr_handler, id_priv);
2383         if (ret)
2384                 goto err;
2385
2386         return 0;
2387 err:
2388         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2389         cma_deref_id(id_priv);
2390         return ret;
2391 }
2392 EXPORT_SYMBOL(rdma_resolve_addr);
2393
2394 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2395 {
2396         struct rdma_id_private *id_priv;
2397         unsigned long flags;
2398         int ret;
2399
2400         id_priv = container_of(id, struct rdma_id_private, id);
2401         spin_lock_irqsave(&id_priv->lock, flags);
2402         if (id_priv->state == RDMA_CM_IDLE) {
2403                 id_priv->reuseaddr = reuse;
2404                 ret = 0;
2405         } else {
2406                 ret = -EINVAL;
2407         }
2408         spin_unlock_irqrestore(&id_priv->lock, flags);
2409         return ret;
2410 }
2411 EXPORT_SYMBOL(rdma_set_reuseaddr);
2412
2413 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2414 {
2415         struct rdma_id_private *id_priv;
2416         unsigned long flags;
2417         int ret;
2418
2419         id_priv = container_of(id, struct rdma_id_private, id);
2420         spin_lock_irqsave(&id_priv->lock, flags);
2421         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2422                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2423                 id_priv->afonly = afonly;
2424                 ret = 0;
2425         } else {
2426                 ret = -EINVAL;
2427         }
2428         spin_unlock_irqrestore(&id_priv->lock, flags);
2429         return ret;
2430 }
2431 EXPORT_SYMBOL(rdma_set_afonly);
2432
2433 static void cma_bind_port(struct rdma_bind_list *bind_list,
2434                           struct rdma_id_private *id_priv)
2435 {
2436         struct sockaddr_in *sin;
2437
2438         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2439         sin->sin_port = htons(bind_list->port);
2440         id_priv->bind_list = bind_list;
2441         hlist_add_head(&id_priv->node, &bind_list->owners);
2442 }
2443
2444 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2445                           unsigned short snum)
2446 {
2447         struct rdma_bind_list *bind_list;
2448         int port, ret;
2449
2450         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2451         if (!bind_list)
2452                 return -ENOMEM;
2453
2454         do {
2455                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2456         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2457
2458         if (ret)
2459                 goto err1;
2460
2461         if (port != snum) {
2462                 ret = -EADDRNOTAVAIL;
2463                 goto err2;
2464         }
2465
2466         bind_list->ps = ps;
2467         bind_list->port = (unsigned short) port;
2468         cma_bind_port(bind_list, id_priv);
2469         return 0;
2470 err2:
2471         idr_remove(ps, port);
2472 err1:
2473         kfree(bind_list);
2474         return ret;
2475 }
2476
2477 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2478 {
2479         static unsigned int last_used_port;
2480         int low, high, remaining;
2481         unsigned int rover;
2482
2483         inet_get_local_port_range(&low, &high);
2484         remaining = (high - low) + 1;
2485         rover = random() % remaining + low;
2486 retry:
2487         if (last_used_port != rover &&
2488             !idr_find(ps, (unsigned short) rover)) {
2489                 int ret = cma_alloc_port(ps, id_priv, rover);
2490                 /*
2491                  * Remember previously used port number in order to avoid
2492                  * re-using same port immediately after it is closed.
2493                  */
2494                 if (!ret)
2495                         last_used_port = rover;
2496                 if (ret != -EADDRNOTAVAIL)
2497                         return ret;
2498                 }
2499         if (--remaining) {
2500                 rover++;
2501                 if ((rover < low) || (rover > high))
2502                         rover = low;
2503                 goto retry;
2504         }
2505         return -EADDRNOTAVAIL;
2506 }
2507
2508 /*
2509  * Check that the requested port is available.  This is called when trying to
2510  * bind to a specific port, or when trying to listen on a bound port.  In
2511  * the latter case, the provided id_priv may already be on the bind_list, but
2512  * we still need to check that it's okay to start listening.
2513  */
2514 static int cma_check_port(struct rdma_bind_list *bind_list,
2515                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2516 {
2517         struct rdma_id_private *cur_id;
2518         struct sockaddr *addr, *cur_addr;
2519
2520         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2521         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2522                 if (id_priv == cur_id)
2523                         continue;
2524
2525                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2526                     cur_id->reuseaddr)
2527                         continue;
2528
2529                 cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2530                 if (id_priv->afonly && cur_id->afonly &&
2531                     (addr->sa_family != cur_addr->sa_family))
2532                         continue;
2533
2534                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2535                         return -EADDRNOTAVAIL;
2536
2537                 if (!cma_addr_cmp(addr, cur_addr))
2538                         return -EADDRINUSE;
2539         }
2540         return 0;
2541 }
2542
2543 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2544 {
2545         struct rdma_bind_list *bind_list;
2546         unsigned short snum;
2547         int ret;
2548
2549         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2550
2551         bind_list = idr_find(ps, snum);
2552         if (!bind_list) {
2553                 ret = cma_alloc_port(ps, id_priv, snum);
2554         } else {
2555                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2556                 if (!ret)
2557                         cma_bind_port(bind_list, id_priv);
2558         }
2559         return ret;
2560 }
2561
2562 static int cma_bind_listen(struct rdma_id_private *id_priv)
2563 {
2564         struct rdma_bind_list *bind_list = id_priv->bind_list;
2565         int ret = 0;
2566
2567         mutex_lock(&lock);
2568         if (bind_list->owners.first->next)
2569                 ret = cma_check_port(bind_list, id_priv, 0);
2570         mutex_unlock(&lock);
2571         return ret;
2572 }
2573
2574 static int cma_get_tcp_port(struct rdma_id_private *id_priv)
2575 {
2576         int ret;
2577         int size;
2578         struct socket *sock;
2579
2580         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2581         if (ret)
2582                 return ret;
2583 #ifdef __linux__
2584         ret = sock->ops->bind(sock,
2585                         (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2586                         ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2587 #else
2588         ret = -sobind(sock,
2589                         (struct sockaddr *)&id_priv->id.route.addr.src_addr,
2590                         curthread);
2591 #endif
2592         if (ret) {
2593                 sock_release(sock);
2594                 return ret;
2595         }
2596
2597         size = ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr);
2598         ret = sock_getname(sock,
2599                         (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2600                         &size, 0);
2601         if (ret) {
2602                 sock_release(sock);
2603                 return ret;
2604         }
2605
2606         id_priv->sock = sock;
2607         return 0;
2608 }
2609
2610 static int cma_get_port(struct rdma_id_private *id_priv)
2611 {
2612         struct idr *ps;
2613         int ret;
2614
2615         switch (id_priv->id.ps) {
2616         case RDMA_PS_SDP:
2617                 ps = &sdp_ps;
2618                 break;
2619         case RDMA_PS_TCP:
2620                 ps = &tcp_ps;
2621                 if (unify_tcp_port_space) {
2622                         ret = cma_get_tcp_port(id_priv);
2623                         if (ret)
2624                                 goto out;
2625                 }
2626                 break;
2627         case RDMA_PS_UDP:
2628                 ps = &udp_ps;
2629                 break;
2630         case RDMA_PS_IPOIB:
2631                 ps = &ipoib_ps;
2632                 break;
2633         case RDMA_PS_IB:
2634                 ps = &ib_ps;
2635                 break;
2636         default:
2637                 return -EPROTONOSUPPORT;
2638         }
2639
2640         mutex_lock(&lock);
2641         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2642                 ret = cma_alloc_any_port(ps, id_priv);
2643         else
2644                 ret = cma_use_port(ps, id_priv);
2645         mutex_unlock(&lock);
2646 out:
2647         return ret;
2648 }
2649
2650 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2651                                struct sockaddr *addr)
2652 {
2653 #if defined(INET6)
2654         struct sockaddr_in6 *sin6;
2655
2656         if (addr->sa_family != AF_INET6)
2657                 return 0;
2658
2659         sin6 = (struct sockaddr_in6 *) addr;
2660         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr) &&
2661             !sin6->sin6_scope_id)
2662                         return -EINVAL;
2663
2664         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2665 #endif
2666         return 0;
2667 }
2668
2669 int rdma_listen(struct rdma_cm_id *id, int backlog)
2670 {
2671         struct rdma_id_private *id_priv;
2672         int ret;
2673
2674         id_priv = container_of(id, struct rdma_id_private, id);
2675         if (id_priv->state == RDMA_CM_IDLE) {
2676                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2677                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2678                 if (ret)
2679                         return ret;
2680         }
2681
2682         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2683                 return -EINVAL;
2684
2685         if (id_priv->reuseaddr) {
2686                 ret = cma_bind_listen(id_priv);
2687                 if (ret)
2688                         goto err;
2689         }
2690
2691         id_priv->backlog = backlog;
2692         if (id->device) {
2693                 switch (rdma_node_get_transport(id->device->node_type)) {
2694                 case RDMA_TRANSPORT_IB:
2695                         ret = cma_ib_listen(id_priv);
2696                         if (ret)
2697                                 goto err;
2698                         break;
2699                 case RDMA_TRANSPORT_IWARP:
2700                 case RDMA_TRANSPORT_SCIF:
2701                         ret = cma_iw_listen(id_priv, backlog);
2702                         if (ret)
2703                                 goto err;
2704                         break;
2705                 default:
2706                         ret = -ENOSYS;
2707                         goto err;
2708                 }
2709         } else
2710                 cma_listen_on_all(id_priv);
2711
2712         return 0;
2713 err:
2714         id_priv->backlog = 0;
2715         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2716         return ret;
2717 }
2718 EXPORT_SYMBOL(rdma_listen);
2719
2720 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2721 {
2722         struct rdma_id_private *id_priv;
2723         int ret;
2724 #if defined(INET6)
2725         int ipv6only;
2726         size_t var_size = sizeof(int);
2727 #endif
2728
2729         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2730                 return -EAFNOSUPPORT;
2731
2732         id_priv = container_of(id, struct rdma_id_private, id);
2733         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2734                 return -EINVAL;
2735
2736         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2737         if (ret)
2738                 goto err1;
2739
2740         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2741         if (!cma_any_addr(addr)) {
2742                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr, NULL);
2743                 if (ret)
2744                         goto err1;
2745
2746                 ret = cma_acquire_dev(id_priv);
2747                 if (ret)
2748                         goto err1;
2749         }
2750
2751         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2752                 if (addr->sa_family == AF_INET)
2753                         id_priv->afonly = 1;
2754 #if defined(INET6)
2755                 else if (addr->sa_family == AF_INET6)
2756                         id_priv->afonly = kernel_sysctlbyname(&thread0, "net.inet6.ip6.v6only",
2757                                             &ipv6only, &var_size, NULL, 0, NULL, 0);
2758 #endif
2759         }
2760         ret = cma_get_port(id_priv);
2761         if (ret)
2762                 goto err2;
2763
2764         return 0;
2765 err2:
2766         if (id_priv->cma_dev)
2767                 cma_release_dev(id_priv);
2768 err1:
2769         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2770         return ret;
2771 }
2772 EXPORT_SYMBOL(rdma_bind_addr);
2773
2774 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2775                           struct rdma_route *route)
2776 {
2777         struct cma_hdr *cma_hdr;
2778         struct sdp_hh *sdp_hdr;
2779
2780         if (route->addr.src_addr.ss_family == AF_INET) {
2781                 struct sockaddr_in *src4, *dst4;
2782
2783                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2784                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2785
2786                 switch (ps) {
2787                 case RDMA_PS_SDP:
2788                         sdp_hdr = hdr;
2789                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2790                                 return -EINVAL;
2791                         sdp_set_ip_ver(sdp_hdr, 4);
2792                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2793                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2794                         sdp_hdr->port = src4->sin_port;
2795                         break;
2796                 default:
2797                         cma_hdr = hdr;
2798                         cma_hdr->cma_version = CMA_VERSION;
2799                         cma_set_ip_ver(cma_hdr, 4);
2800                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2801                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2802                         cma_hdr->port = src4->sin_port;
2803                         break;
2804                 }
2805         } else {
2806                 struct sockaddr_in6 *src6, *dst6;
2807
2808                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2809                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2810
2811                 switch (ps) {
2812                 case RDMA_PS_SDP:
2813                         sdp_hdr = hdr;
2814                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2815                                 return -EINVAL;
2816                         sdp_set_ip_ver(sdp_hdr, 6);
2817                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2818                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2819                         sdp_hdr->port = src6->sin6_port;
2820                         break;
2821                 default:
2822                         cma_hdr = hdr;
2823                         cma_hdr->cma_version = CMA_VERSION;
2824                         cma_set_ip_ver(cma_hdr, 6);
2825                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2826                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2827                         cma_hdr->port = src6->sin6_port;
2828                         break;
2829                 }
2830         }
2831         return 0;
2832 }
2833
2834 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2835                                 struct ib_cm_event *ib_event)
2836 {
2837         struct rdma_id_private *id_priv = cm_id->context;
2838         struct rdma_cm_event event;
2839         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2840         int ret = 0;
2841
2842         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2843                 return 0;
2844
2845         memset(&event, 0, sizeof event);
2846         switch (ib_event->event) {
2847         case IB_CM_SIDR_REQ_ERROR:
2848                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2849                 event.status = -ETIMEDOUT;
2850                 break;
2851         case IB_CM_SIDR_REP_RECEIVED:
2852                 event.param.ud.private_data = ib_event->private_data;
2853                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2854                 if (rep->status != IB_SIDR_SUCCESS) {
2855                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2856                         event.status = ib_event->param.sidr_rep_rcvd.status;
2857                         break;
2858                 }
2859                 ret = cma_set_qkey(id_priv);
2860                 if (ret) {
2861                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2862                         event.status = -EINVAL;
2863                         break;
2864                 }
2865                 if (id_priv->qkey != rep->qkey) {
2866                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2867                         event.status = -EINVAL;
2868                         break;
2869                 }
2870                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2871                                      id_priv->id.route.path_rec,
2872                                      &event.param.ud.ah_attr);
2873                 event.param.ud.qp_num = rep->qpn;
2874                 event.param.ud.qkey = rep->qkey;
2875                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2876                 event.status = 0;
2877                 break;
2878         default:
2879                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2880                        ib_event->event);
2881                 goto out;
2882         }
2883
2884         ret = id_priv->id.event_handler(&id_priv->id, &event);
2885         if (ret) {
2886                 /* Destroy the CM ID by returning a non-zero value. */
2887                 id_priv->cm_id.ib = NULL;
2888                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2889                 mutex_unlock(&id_priv->handler_mutex);
2890                 rdma_destroy_id(&id_priv->id);
2891                 return ret;
2892         }
2893 out:
2894         mutex_unlock(&id_priv->handler_mutex);
2895         return ret;
2896 }
2897
2898 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2899                               struct rdma_conn_param *conn_param)
2900 {
2901         struct ib_cm_sidr_req_param req;
2902         struct rdma_route *route;
2903         struct ib_cm_id *id;
2904         int ret;
2905
2906         req.private_data_len = sizeof(struct cma_hdr) +
2907                                conn_param->private_data_len;
2908         if (req.private_data_len < conn_param->private_data_len)
2909                 return -EINVAL;
2910
2911         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2912         if (!req.private_data)
2913                 return -ENOMEM;
2914
2915         if (conn_param->private_data && conn_param->private_data_len)
2916                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2917                        conn_param->private_data, conn_param->private_data_len);
2918
2919         route = &id_priv->id.route;
2920         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2921         if (ret)
2922                 goto out;
2923
2924         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2925                              id_priv);
2926         if (IS_ERR(id)) {
2927                 ret = PTR_ERR(id);
2928                 goto out;
2929         }
2930         id_priv->cm_id.ib = id;
2931
2932         req.path = route->path_rec;
2933         req.service_id = cma_get_service_id(id_priv->id.ps,
2934                                             (struct sockaddr *) &route->addr.dst_addr);
2935         req.timeout_ms = 1 << (cma_response_timeout - 8);
2936         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2937
2938         cma_dbg(id_priv, "sending SIDR\n");
2939         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2940         if (ret) {
2941                 ib_destroy_cm_id(id_priv->cm_id.ib);
2942                 id_priv->cm_id.ib = NULL;
2943         }
2944 out:
2945         kfree(req.private_data);
2946         return ret;
2947 }
2948
2949 static int cma_connect_ib(struct rdma_id_private *id_priv,
2950                           struct rdma_conn_param *conn_param)
2951 {
2952         struct ib_cm_req_param req;
2953         struct rdma_route *route;
2954         void *private_data;
2955         struct ib_cm_id *id;
2956         int offset, ret;
2957
2958         memset(&req, 0, sizeof req);
2959         offset = cma_user_data_offset(id_priv->id.ps);
2960         req.private_data_len = offset + conn_param->private_data_len;
2961         if (req.private_data_len < conn_param->private_data_len)
2962                 return -EINVAL;
2963
2964         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2965         if (!private_data)
2966                 return -ENOMEM;
2967
2968         if (conn_param->private_data && conn_param->private_data_len)
2969                 memcpy(private_data + offset, conn_param->private_data,
2970                        conn_param->private_data_len);
2971
2972         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2973         if (IS_ERR(id)) {
2974                 ret = PTR_ERR(id);
2975                 goto out;
2976         }
2977         id_priv->cm_id.ib = id;
2978
2979         route = &id_priv->id.route;
2980         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2981         if (ret)
2982                 goto out;
2983         req.private_data = private_data;
2984
2985         req.primary_path = &route->path_rec[0];
2986         if (route->num_paths == 2)
2987                 req.alternate_path = &route->path_rec[1];
2988
2989         req.service_id = cma_get_service_id(id_priv->id.ps,
2990                                             (struct sockaddr *) &route->addr.dst_addr);
2991         req.qp_num = id_priv->qp_num;
2992         req.qp_type = id_priv->id.qp_type;
2993         req.starting_psn = id_priv->seq_num;
2994         req.responder_resources = conn_param->responder_resources;
2995         req.initiator_depth = conn_param->initiator_depth;
2996         req.flow_control = conn_param->flow_control;
2997         req.retry_count = min_t(u8, 7, conn_param->retry_count);
2998         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2999        req.remote_cm_response_timeout = cma_response_timeout;
3000        req.local_cm_response_timeout = cma_response_timeout;
3001         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3002         req.srq = id_priv->srq ? 1 : 0;
3003
3004         cma_dbg(id_priv, "sending REQ\n");
3005         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3006 out:
3007         if (ret && !IS_ERR(id)) {
3008                 ib_destroy_cm_id(id);
3009                 id_priv->cm_id.ib = NULL;
3010         }
3011
3012         kfree(private_data);
3013         return ret;
3014 }
3015
3016 static int cma_connect_iw(struct rdma_id_private *id_priv,
3017                           struct rdma_conn_param *conn_param)
3018 {
3019         struct iw_cm_id *cm_id;
3020         struct sockaddr_in* sin;
3021         int ret;
3022         struct iw_cm_conn_param iw_param;
3023
3024         cm_id = iw_create_cm_id(id_priv->id.device, id_priv->sock,
3025                                         cma_iw_handler, id_priv);
3026         if (IS_ERR(cm_id))
3027                 return PTR_ERR(cm_id);
3028
3029         id_priv->cm_id.iw = cm_id;
3030
3031         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
3032         cm_id->local_addr = *sin;
3033
3034         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
3035         cm_id->remote_addr = *sin;
3036
3037         ret = cma_modify_qp_rtr(id_priv, conn_param);
3038         if (ret)
3039                 goto out;
3040
3041         if (conn_param) {
3042         iw_param.ord = conn_param->initiator_depth;
3043         iw_param.ird = conn_param->responder_resources;
3044         iw_param.private_data = conn_param->private_data;
3045         iw_param.private_data_len = conn_param->private_data_len;
3046                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3047         } else {
3048                 memset(&iw_param, 0, sizeof iw_param);
3049                 iw_param.qpn = id_priv->qp_num;
3050         }
3051         ret = iw_cm_connect(cm_id, &iw_param);
3052 out:
3053         if (ret) {
3054                 iw_destroy_cm_id(cm_id);
3055                 id_priv->cm_id.iw = NULL;
3056         }
3057         return ret;
3058 }
3059
3060 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3061 {
3062         struct rdma_id_private *id_priv;
3063         int ret;
3064
3065         id_priv = container_of(id, struct rdma_id_private, id);
3066         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3067                 return -EINVAL;
3068
3069         if (!id->qp) {
3070                 id_priv->qp_num = conn_param->qp_num;
3071                 id_priv->srq = conn_param->srq;
3072         }
3073
3074         switch (rdma_node_get_transport(id->device->node_type)) {
3075         case RDMA_TRANSPORT_IB:
3076                 if (id->qp_type == IB_QPT_UD)
3077                         ret = cma_resolve_ib_udp(id_priv, conn_param);
3078                 else
3079                         ret = cma_connect_ib(id_priv, conn_param);
3080                 break;
3081         case RDMA_TRANSPORT_IWARP:
3082         case RDMA_TRANSPORT_SCIF:
3083                 ret = cma_connect_iw(id_priv, conn_param);
3084                 break;
3085         default:
3086                 ret = -ENOSYS;
3087                 break;
3088         }
3089         if (ret)
3090                 goto err;
3091
3092         return 0;
3093 err:
3094         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3095         return ret;
3096 }
3097 EXPORT_SYMBOL(rdma_connect);
3098
3099 static int cma_accept_ib(struct rdma_id_private *id_priv,
3100                          struct rdma_conn_param *conn_param)
3101 {
3102         struct ib_cm_rep_param rep;
3103         int ret;
3104
3105         ret = cma_modify_qp_rtr(id_priv, conn_param);
3106         if (ret)
3107                 goto out;
3108
3109         ret = cma_modify_qp_rts(id_priv, conn_param);
3110         if (ret)
3111                 goto out;
3112
3113         memset(&rep, 0, sizeof rep);
3114         rep.qp_num = id_priv->qp_num;
3115         rep.starting_psn = id_priv->seq_num;
3116         rep.private_data = conn_param->private_data;
3117         rep.private_data_len = conn_param->private_data_len;
3118         rep.responder_resources = conn_param->responder_resources;
3119         rep.initiator_depth = conn_param->initiator_depth;
3120         rep.failover_accepted = 0;
3121         rep.flow_control = conn_param->flow_control;
3122         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3123         rep.srq = id_priv->srq ? 1 : 0;
3124         cma_dbg(id_priv, "sending REP\n");
3125         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3126 out:
3127         return ret;
3128 }
3129
3130 static int cma_accept_iw(struct rdma_id_private *id_priv,
3131                   struct rdma_conn_param *conn_param)
3132 {
3133         struct iw_cm_conn_param iw_param;
3134         int ret;
3135
3136         if (!conn_param)
3137                 return -EINVAL;
3138
3139         ret = cma_modify_qp_rtr(id_priv, conn_param);
3140         if (ret)
3141                 return ret;
3142
3143         iw_param.ord = conn_param->initiator_depth;
3144         iw_param.ird = conn_param->responder_resources;
3145         iw_param.private_data = conn_param->private_data;
3146         iw_param.private_data_len = conn_param->private_data_len;
3147         if (id_priv->id.qp) {
3148                 iw_param.qpn = id_priv->qp_num;
3149         } else
3150                 iw_param.qpn = conn_param->qp_num;
3151
3152         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3153 }
3154
3155 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3156                              enum ib_cm_sidr_status status,
3157                              const void *private_data, int private_data_len)
3158 {
3159         struct ib_cm_sidr_rep_param rep;
3160         int ret;
3161
3162         memset(&rep, 0, sizeof rep);
3163         rep.status = status;
3164         if (status == IB_SIDR_SUCCESS) {
3165                 ret = cma_set_qkey(id_priv);
3166                 if (ret)
3167                         return ret;
3168                 rep.qp_num = id_priv->qp_num;
3169                 rep.qkey = id_priv->qkey;
3170         }
3171         rep.private_data = private_data;
3172         rep.private_data_len = private_data_len;
3173
3174         cma_dbg(id_priv, "sending SIDR\n");
3175         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3176 }
3177
3178 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3179 {
3180         struct rdma_id_private *id_priv;
3181         int ret;
3182
3183         id_priv = container_of(id, struct rdma_id_private, id);
3184
3185         id_priv->owner = curthread->td_proc->p_pid;
3186         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3187                 return -EINVAL;
3188
3189         if (!id->qp && conn_param) {
3190                 id_priv->qp_num = conn_param->qp_num;
3191                 id_priv->srq = conn_param->srq;
3192         }
3193
3194         switch (rdma_node_get_transport(id->device->node_type)) {
3195         case RDMA_TRANSPORT_IB:
3196                 if (id->qp_type == IB_QPT_UD) {
3197                         if (conn_param)
3198                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3199                                                 conn_param->private_data,
3200                                                 conn_param->private_data_len);
3201                         else
3202                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3203                                                         NULL, 0);
3204                 } else {
3205                         if (conn_param)
3206                         ret = cma_accept_ib(id_priv, conn_param);
3207                 else
3208                         ret = cma_rep_recv(id_priv);
3209                 }
3210                 break;
3211         case RDMA_TRANSPORT_IWARP:
3212         case RDMA_TRANSPORT_SCIF:
3213                 ret = cma_accept_iw(id_priv, conn_param);
3214                 break;
3215         default:
3216                 ret = -ENOSYS;
3217                 break;
3218         }
3219
3220         if (ret)
3221                 goto reject;
3222
3223         return 0;
3224 reject:
3225         cma_modify_qp_err(id_priv);
3226         rdma_reject(id, NULL, 0);
3227         return ret;
3228 }
3229 EXPORT_SYMBOL(rdma_accept);
3230
3231 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3232 {
3233         struct rdma_id_private *id_priv;
3234         int ret;
3235
3236         id_priv = container_of(id, struct rdma_id_private, id);
3237         if (!id_priv->cm_id.ib)
3238                 return -EINVAL;
3239
3240         switch (id->device->node_type) {
3241         case RDMA_NODE_IB_CA:
3242                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3243                 break;
3244         default:
3245                 ret = 0;
3246                 break;
3247         }
3248         return ret;
3249 }
3250 EXPORT_SYMBOL(rdma_notify);
3251
3252 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3253                 u8 private_data_len)
3254 {
3255         struct rdma_id_private *id_priv;
3256         int ret;
3257
3258         id_priv = container_of(id, struct rdma_id_private, id);
3259         if (!id_priv->cm_id.ib)
3260                 return -EINVAL;
3261
3262         switch (rdma_node_get_transport(id->device->node_type)) {
3263         case RDMA_TRANSPORT_IB:
3264                 if (id->qp_type == IB_QPT_UD)
3265                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
3266                                                 private_data, private_data_len);
3267                 else {
3268                         cma_dbg(id_priv, "sending REJ\n");
3269                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3270                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3271                                              0, private_data, private_data_len);
3272                 }
3273                 break;
3274         case RDMA_TRANSPORT_IWARP:
3275         case RDMA_TRANSPORT_SCIF:
3276                 ret = iw_cm_reject(id_priv->cm_id.iw,
3277                                    private_data, private_data_len);
3278                 break;
3279         default:
3280                 ret = -ENOSYS;
3281                 break;
3282         }
3283         return ret;
3284 }
3285 EXPORT_SYMBOL(rdma_reject);
3286
3287 int rdma_disconnect(struct rdma_cm_id *id)
3288 {
3289         struct rdma_id_private *id_priv;
3290         int ret;
3291
3292         id_priv = container_of(id, struct rdma_id_private, id);
3293         if (!id_priv->cm_id.ib)
3294                 return -EINVAL;
3295
3296         switch (rdma_node_get_transport(id->device->node_type)) {
3297         case RDMA_TRANSPORT_IB:
3298                 ret = cma_modify_qp_err(id_priv);
3299                 if (ret)
3300                         goto out;
3301                 /* Initiate or respond to a disconnect. */
3302                 cma_dbg(id_priv, "sending DREQ\n");
3303                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
3304                         cma_dbg(id_priv, "sending DREP\n");
3305                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3306                 }
3307                 break;
3308         case RDMA_TRANSPORT_IWARP:
3309         case RDMA_TRANSPORT_SCIF:
3310                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3311                 break;
3312         default:
3313                 ret = -EINVAL;
3314                 break;
3315         }
3316 out:
3317         return ret;
3318 }
3319 EXPORT_SYMBOL(rdma_disconnect);
3320
3321 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3322 {
3323         struct rdma_id_private *id_priv;
3324         struct cma_multicast *mc = multicast->context;
3325         struct rdma_cm_event event;
3326         struct rdma_dev_addr *dev_addr;
3327         int ret;
3328         struct net_device *ndev = NULL;
3329         u16 vlan;
3330
3331         id_priv = mc->id_priv;
3332         dev_addr = &id_priv->id.route.addr.dev_addr;
3333         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3334             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3335                 return 0;
3336
3337         mutex_lock(&id_priv->qp_mutex);
3338         if (!status && id_priv->id.qp)
3339                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3340                                          be16_to_cpu(multicast->rec.mlid));
3341         mutex_unlock(&id_priv->qp_mutex);
3342
3343         memset(&event, 0, sizeof event);
3344         event.status = status;
3345         event.param.ud.private_data = mc->context;
3346         ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3347         if (!ndev) {
3348                 status = -ENODEV;
3349         } else {
3350                 vlan = rdma_vlan_dev_vlan_id(ndev);
3351                 dev_put(ndev);
3352         }
3353         if (!status) {
3354                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3355                 ib_init_ah_from_mcmember(id_priv->id.device,
3356                                          id_priv->id.port_num, &multicast->rec,
3357                                          &event.param.ud.ah_attr);
3358                 event.param.ud.ah_attr.vlan_id = vlan;
3359                 event.param.ud.qp_num = 0xFFFFFF;
3360                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3361         } else {
3362                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3363
3364                 /* mark that the cached record is no longer valid */
3365                 if (status != -ENETRESET && status != -EAGAIN) {
3366                         spin_lock(&id_priv->lock);
3367                         id_priv->is_valid_rec = 0;
3368                         spin_unlock(&id_priv->lock);
3369                 }
3370         }
3371
3372         ret = id_priv->id.event_handler(&id_priv->id, &event);
3373         if (ret) {
3374                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3375                 mutex_unlock(&id_priv->handler_mutex);
3376                 rdma_destroy_id(&id_priv->id);
3377                 return 0;
3378         }
3379
3380         mutex_unlock(&id_priv->handler_mutex);
3381         return 0;
3382 }
3383
3384 static void cma_set_mgid(struct rdma_id_private *id_priv,
3385                          struct sockaddr *addr, union ib_gid *mgid)
3386 {
3387         unsigned char mc_map[MAX_ADDR_LEN];
3388         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3389         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3390 #if defined(INET6)
3391         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3392 #endif
3393
3394         if (cma_any_addr(addr)) {
3395                 memset(mgid, 0, sizeof *mgid);
3396 #if defined(INET6)
3397         } else if ((addr->sa_family == AF_INET6) &&
3398                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3399                                                                  0xFF10A01B)) {
3400                 /* IPv6 address is an SA assigned MGID. */
3401                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3402         } else if (addr->sa_family == AF_INET6) {
3403                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3404                 if (id_priv->id.ps == RDMA_PS_UDP)
3405                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3406                 *mgid = *(union ib_gid *) (mc_map + 4);
3407 #endif
3408         } else {
3409                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3410                 if (id_priv->id.ps == RDMA_PS_UDP)
3411                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3412                 *mgid = *(union ib_gid *) (mc_map + 4);
3413         }
3414 }
3415
3416 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3417                                  struct cma_multicast *mc)
3418 {
3419         struct ib_sa_mcmember_rec rec;
3420         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3421         ib_sa_comp_mask comp_mask;
3422         int ret = 0;
3423
3424         ib_addr_get_mgid(dev_addr, &id_priv->rec.mgid);
3425
3426         /* cache ipoib bc record */
3427         spin_lock(&id_priv->lock);
3428         if (!id_priv->is_valid_rec)
3429                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
3430                                              id_priv->id.port_num,
3431                                              &id_priv->rec.mgid,
3432                                              &id_priv->rec);
3433         if (ret) {
3434                 id_priv->is_valid_rec = 0;
3435                 spin_unlock(&id_priv->lock);
3436                 return ret;
3437         } else {
3438                 rec = id_priv->rec;
3439                 id_priv->is_valid_rec = 1;
3440         }
3441         spin_unlock(&id_priv->lock);
3442
3443         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3444         if (id_priv->id.ps == RDMA_PS_UDP)
3445                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3446         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3447         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3448         rec.join_state = 1;
3449
3450         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3451                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3452                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3453                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3454                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3455
3456         if (id_priv->id.ps == RDMA_PS_IPOIB)
3457                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3458                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3459                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3460                              IB_SA_MCMEMBER_REC_MTU |
3461                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3462
3463         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3464                                                 id_priv->id.port_num, &rec,
3465                                                 comp_mask, GFP_KERNEL,
3466                                                 cma_ib_mc_handler, mc);
3467         return PTR_RET(mc->multicast.ib);
3468 }
3469
3470 static void iboe_mcast_work_handler(struct work_struct *work)
3471 {
3472         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3473         struct cma_multicast *mc = mw->mc;
3474         struct ib_sa_multicast *m = mc->multicast.ib;
3475
3476         mc->multicast.ib->context = mc;
3477         cma_ib_mc_handler(0, m);
3478         kref_put(&mc->mcref, release_mc);
3479         kfree(mw);
3480 }
3481
3482 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3483 {
3484         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3485         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3486
3487         if (cma_any_addr(addr)) {
3488                 memset(mgid, 0, sizeof *mgid);
3489         } else if (addr->sa_family == AF_INET6) {
3490                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3491         } else {
3492                 mgid->raw[0] = 0xff;
3493                 mgid->raw[1] = 0x0e;
3494                 mgid->raw[2] = 0;
3495                 mgid->raw[3] = 0;
3496                 mgid->raw[4] = 0;
3497                 mgid->raw[5] = 0;
3498                 mgid->raw[6] = 0;
3499                 mgid->raw[7] = 0;
3500                 mgid->raw[8] = 0;
3501                 mgid->raw[9] = 0;
3502                 mgid->raw[10] = 0xff;
3503                 mgid->raw[11] = 0xff;
3504                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3505         }
3506 }
3507
3508 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3509                                    struct cma_multicast *mc)
3510 {
3511         struct iboe_mcast_work *work;
3512         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3513         int err;
3514         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3515         struct net_device *ndev = NULL;
3516
3517         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3518                 return -EINVAL;
3519
3520         work = kzalloc(sizeof *work, GFP_KERNEL);
3521         if (!work)
3522                 return -ENOMEM;
3523
3524         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3525         if (!mc->multicast.ib) {
3526                 err = -ENOMEM;
3527                 goto out1;
3528         }
3529
3530         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3531
3532         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3533         if (id_priv->id.ps == RDMA_PS_UDP)
3534                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3535
3536         if (dev_addr->bound_dev_if)
3537                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3538         if (!ndev) {
3539                 err = -ENODEV;
3540                 goto out2;
3541         }
3542         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3543         mc->multicast.ib->rec.hop_limit = 1;
3544         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->if_mtu);
3545         dev_put(ndev);
3546         if (!mc->multicast.ib->rec.mtu) {
3547                 err = -EINVAL;
3548                 goto out2;
3549         }
3550         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3551                     &mc->multicast.ib->rec.port_gid);
3552         work->id = id_priv;
3553         work->mc = mc;
3554         INIT_WORK(&work->work, iboe_mcast_work_handler);
3555         kref_get(&mc->mcref);
3556         queue_work(cma_wq, &work->work);
3557
3558         return 0;
3559
3560 out2:
3561         kfree(mc->multicast.ib);
3562 out1:
3563         kfree(work);
3564         return err;
3565 }
3566
3567 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3568                         void *context)
3569 {
3570         struct rdma_id_private *id_priv;
3571         struct cma_multicast *mc;
3572         int ret;
3573
3574         id_priv = container_of(id, struct rdma_id_private, id);
3575         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3576             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3577                 return -EINVAL;
3578
3579         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3580         if (!mc)
3581                 return -ENOMEM;
3582
3583         memcpy(&mc->addr, addr, ip_addr_size(addr));
3584         mc->context = context;
3585         mc->id_priv = id_priv;
3586
3587         spin_lock(&id_priv->lock);
3588         list_add(&mc->list, &id_priv->mc_list);
3589         spin_unlock(&id_priv->lock);
3590
3591         switch (rdma_node_get_transport(id->device->node_type)) {
3592         case RDMA_TRANSPORT_IB:
3593                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3594                 case IB_LINK_LAYER_INFINIBAND:
3595                         ret = cma_join_ib_multicast(id_priv, mc);
3596                         break;
3597                 case IB_LINK_LAYER_ETHERNET:
3598                         kref_init(&mc->mcref);
3599                         ret = cma_iboe_join_multicast(id_priv, mc);
3600                         break;
3601                 default:
3602                         ret = -EINVAL;
3603                 }
3604                 break;
3605         default:
3606                 ret = -ENOSYS;
3607                 break;
3608         }
3609
3610         if (ret) {
3611                 spin_lock_irq(&id_priv->lock);
3612                 list_del(&mc->list);
3613                 spin_unlock_irq(&id_priv->lock);
3614                 kfree(mc);
3615         }
3616         return ret;
3617 }
3618 EXPORT_SYMBOL(rdma_join_multicast);
3619
3620 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3621 {
3622         struct rdma_id_private *id_priv;
3623         struct cma_multicast *mc;
3624
3625         id_priv = container_of(id, struct rdma_id_private, id);
3626         spin_lock_irq(&id_priv->lock);
3627         list_for_each_entry(mc, &id_priv->mc_list, list) {
3628                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3629                         list_del(&mc->list);
3630                         spin_unlock_irq(&id_priv->lock);
3631
3632                         if (id->qp)
3633                                 ib_detach_mcast(id->qp,
3634                                                 &mc->multicast.ib->rec.mgid,
3635                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3636                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3637                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3638                                 case IB_LINK_LAYER_INFINIBAND:
3639                                         ib_sa_free_multicast(mc->multicast.ib);
3640                                         kfree(mc);
3641                                         break;
3642                                 case IB_LINK_LAYER_ETHERNET:
3643                                         kref_put(&mc->mcref, release_mc);
3644                                         break;
3645                                 default:
3646                                         break;
3647                                 }
3648                         }
3649                         return;
3650                 }
3651         }
3652         spin_unlock_irq(&id_priv->lock);
3653 }
3654 EXPORT_SYMBOL(rdma_leave_multicast);
3655
3656 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3657 {
3658         struct rdma_dev_addr *dev_addr;
3659         struct cma_ndev_work *work;
3660
3661         dev_addr = &id_priv->id.route.addr.dev_addr;
3662
3663         if ((dev_addr->bound_dev_if == ndev->if_index) &&
3664             memcmp(dev_addr->src_dev_addr, IF_LLADDR(ndev), ndev->if_addrlen)) {
3665                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3666                        ndev->if_xname, &id_priv->id);
3667                 work = kzalloc(sizeof *work, GFP_KERNEL);
3668                 if (!work)
3669                         return -ENOMEM;
3670
3671                 INIT_WORK(&work->work, cma_ndev_work_handler);
3672                 work->id = id_priv;
3673                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3674                 atomic_inc(&id_priv->refcount);
3675                 queue_work(cma_wq, &work->work);
3676         }
3677
3678         return 0;
3679 }
3680
3681 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3682                                void *ctx)
3683 {
3684         struct net_device *ndev = (struct net_device *)ctx;
3685         struct cma_device *cma_dev;
3686         struct rdma_id_private *id_priv;
3687         int ret = NOTIFY_DONE;
3688
3689 /* BONDING related, commented out until the bonding is resolved */
3690 #if 0
3691         if (dev_net(ndev) != &init_net)
3692                 return NOTIFY_DONE;
3693
3694         if (event != NETDEV_BONDING_FAILOVER)
3695                 return NOTIFY_DONE;
3696
3697         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3698                 return NOTIFY_DONE;
3699 #endif
3700         if (event != NETDEV_DOWN && event != NETDEV_UNREGISTER)
3701                 return NOTIFY_DONE;
3702
3703         mutex_lock(&lock);
3704         list_for_each_entry(cma_dev, &dev_list, list)
3705                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3706                         ret = cma_netdev_change(ndev, id_priv);
3707                         if (ret)
3708                                 goto out;
3709                 }
3710
3711 out:
3712         mutex_unlock(&lock);
3713         return ret;
3714 }
3715
3716 static struct notifier_block cma_nb = {
3717         .notifier_call = cma_netdev_callback
3718 };
3719
3720 static void cma_add_one(struct ib_device *device)
3721 {
3722         struct cma_device *cma_dev;
3723         struct rdma_id_private *id_priv;
3724
3725         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3726         if (!cma_dev)
3727                 return;
3728
3729         cma_dev->device = device;
3730
3731         init_completion(&cma_dev->comp);
3732         atomic_set(&cma_dev->refcount, 1);
3733         INIT_LIST_HEAD(&cma_dev->id_list);
3734         ib_set_client_data(device, &cma_client, cma_dev);
3735
3736         mutex_lock(&lock);
3737         list_add_tail(&cma_dev->list, &dev_list);
3738         list_for_each_entry(id_priv, &listen_any_list, list)
3739                 cma_listen_on_dev(id_priv, cma_dev);
3740         mutex_unlock(&lock);
3741 }
3742
3743 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3744 {
3745         struct rdma_cm_event event;
3746         enum rdma_cm_state state;
3747         int ret = 0;
3748
3749         /* Record that we want to remove the device */
3750         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3751         if (state == RDMA_CM_DESTROYING)
3752                 return 0;
3753
3754         cma_cancel_operation(id_priv, state);
3755         mutex_lock(&id_priv->handler_mutex);
3756
3757         /* Check for destruction from another callback. */
3758         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3759                 goto out;
3760
3761         memset(&event, 0, sizeof event);
3762         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3763         ret = id_priv->id.event_handler(&id_priv->id, &event);
3764 out:
3765         mutex_unlock(&id_priv->handler_mutex);
3766         return ret;
3767 }
3768
3769 static void cma_process_remove(struct cma_device *cma_dev)
3770 {
3771         struct rdma_id_private *id_priv;
3772         int ret;
3773
3774         mutex_lock(&lock);
3775         while (!list_empty(&cma_dev->id_list)) {
3776                 id_priv = list_entry(cma_dev->id_list.next,
3777                                      struct rdma_id_private, list);
3778
3779                 list_del(&id_priv->listen_list);
3780                 list_del_init(&id_priv->list);
3781                 atomic_inc(&id_priv->refcount);
3782                 mutex_unlock(&lock);
3783
3784                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3785                 cma_deref_id(id_priv);
3786                 if (ret)
3787                         rdma_destroy_id(&id_priv->id);
3788
3789                 mutex_lock(&lock);
3790         }
3791         mutex_unlock(&lock);
3792
3793         cma_deref_dev(cma_dev);
3794         wait_for_completion(&cma_dev->comp);
3795 }
3796
3797 static void cma_remove_one(struct ib_device *device)
3798 {
3799         struct cma_device *cma_dev;
3800
3801         cma_dev = ib_get_client_data(device, &cma_client);
3802         if (!cma_dev)
3803                 return;
3804
3805         mutex_lock(&lock);
3806         list_del(&cma_dev->list);
3807         mutex_unlock(&lock);
3808
3809         cma_process_remove(cma_dev);
3810         kfree(cma_dev);
3811 }
3812
3813 static int __init cma_init(void)
3814 {
3815         int ret = -ENOMEM;
3816
3817         cma_wq = create_singlethread_workqueue("rdma_cm");
3818         if (!cma_wq)
3819                 return -ENOMEM;
3820
3821         cma_free_wq = create_singlethread_workqueue("rdma_cm_fr");
3822         if (!cma_free_wq)
3823                 goto err1;
3824
3825         ib_sa_register_client(&sa_client);
3826         rdma_addr_register_client(&addr_client);
3827         register_netdevice_notifier(&cma_nb);
3828
3829         ret = ib_register_client(&cma_client);
3830         if (ret)
3831                 goto err;
3832
3833         return 0;
3834
3835 err:
3836         unregister_netdevice_notifier(&cma_nb);
3837         rdma_addr_unregister_client(&addr_client);
3838         ib_sa_unregister_client(&sa_client);
3839
3840         destroy_workqueue(cma_free_wq);
3841 err1:
3842         destroy_workqueue(cma_wq);
3843         return ret;
3844 }
3845
3846 static void __exit cma_cleanup(void)
3847 {
3848         ib_unregister_client(&cma_client);
3849         unregister_netdevice_notifier(&cma_nb);
3850         rdma_addr_unregister_client(&addr_client);
3851         ib_sa_unregister_client(&sa_client);
3852         flush_workqueue(cma_free_wq);
3853         destroy_workqueue(cma_free_wq);
3854         destroy_workqueue(cma_wq);
3855         idr_destroy(&sdp_ps);
3856         idr_destroy(&tcp_ps);
3857         idr_destroy(&udp_ps);
3858         idr_destroy(&ipoib_ps);
3859         idr_destroy(&ib_ps);
3860 }
3861
3862 module_init(cma_init);
3863 module_exit(cma_cleanup);